Daily Papers

We present a generative dialogue system capable of operating in a full-duplex manner, allowing for seamless interaction. It is based on a large language model (LLM) carefully aligned to be aware of a perception module, a motor function module, and the concept of a simple finite state machine (called neural FSM) with two states. The perception and motor function modules operate in tandem, allowing the system to speak and listen to the user simultaneously. The LLM generates textual tokens for inquiry responses and makes autonomous decisions to start responding to, wait for, or interrupt the user by emitting control tokens to the neural FSM. All these tasks of the LLM are carried out as next token prediction on a serialized view of the dialogue in real-time. In automatic quality evaluations simulating real-life interaction, the proposed system reduces the average conversation response latency by more than threefold compared with LLM-based half-duplex dialogue systems while responding within less than 500 milliseconds in more than 50% of evaluated interactions. Running an LLM with only 8 billion parameters, our system exhibits an 8% higher interruption precision rate than the best available commercial LLM for voice-based dialogue.

Voice agents deployed in structured workflows (customer service, healthcare scheduling, account management) must handle frequent user interruptions while maintaining progress through multi-step procedures. Existing benchmarks for speech-capable models focus on the timing of interruptions: barge-in detection, endpointing, and turn-taking dynamics. They leave unmeasured what happens after the interruption: does the agent resume the workflow at the correct step? Does it address the user's interjection? Does it avoid re-delivering content the user already heard? We introduce IHBench (Interruption Handling Benchmark), a benchmark that evaluates post-interruption recovery in voice agents executing state-machine-driven workflows across 10 enterprise domains. Six interruption types are injected at controlled points mid-utterance, with per-interruption evaluation rubrics generated alongside the data. Each interruption is scored on two axes: task fulfillment and recovery quality. We evaluate 27 audio-language model configurations from OpenAI, Google, and the open-weight community. Models vary widely, and recovery quality depends strongly on the interruption type. Across our experiments, closed-weight models are consistently more robust to interruptions than open-weight ones: they win far more often on task fulfillment, degrade roughly 3.3x more slowly as conversations grow longer, and show no audio-versus-text modality gap, whereas the open-weight models lose ground on all three. A human study validates the LLM judge against human annotators, and a cross-benchmark analysis against AudioMultiChallenge indicates that recovery quality is a largely distinct capability axis.

Full-duplex speech models listen and speak at once, promising always-on assistants. Yet they must also decide when they should speak. Human listeners speak when addressed or when the speaker stops, but also self-select to correct a false claim, supply a missing word, or warn of danger. We ask whether full-duplex models do the same. To separate the reason to speak from the opportunity, we construct context-matched English monologues in which only the trigger utterance varies within a topic, define 10 conditions from turn-allocation rules, and compress inter-word pauses to limit opportunities created by silence. Across five model families, being addressed and silence are far more reliable triggers than false facts or hazards. Frame-level text-token probabilities in Moshi and PersonaPlex are lower for false facts than for Neutral when averaged over the first 2\,s after trigger end. Pauses or permission to interrupt do not close this gap either. Given the floor, Moshi and PersonaPlex answer most direct questions, yet the proportion of non-empty false-fact replies that challenge the claim is only .14--.15, and the proportion of hazard replies that warn of danger is .04--.07. This paper thus identifies a gap in both speech initiation and response content. Closing it requires genuine content understanding and intervention decisions grounded in it.

Voice dictation is an increasingly important text input modality. Existing systems that allow both dictation and editing-by-voice restrict their command language to flat templates invoked by trigger words. In this work, we study the feasibility of allowing users to interrupt their dictation with spoken editing commands in open-ended natural language. We introduce a new task and dataset, TERTiUS, to experiment with such systems. To support this flexibility in real-time, a system must incrementally segment and classify spans of speech as either dictation or command, and interpret the spans that are commands. We experiment with using large pre-trained language models to predict the edited text, or alternatively, to predict a small text-editing program. Experiments show a natural trade-off between model accuracy and latency: a smaller model achieves 30% end-state accuracy with 1.3 seconds of latency, while a larger model achieves 55% end-state accuracy with 7 seconds of latency.

Achieving natural full-duplex interaction in spoken dialogue systems (SDS) remains a challenge due to the difficulty of accurately detecting user interruptions. Current solutions are polarized between "trigger-happy" VAD-based methods that misinterpret backchannels and robust end-to-end models that exhibit unacceptable response delays. Moreover, the absence of real-world benchmarks and holistic metrics hinders progress in the field. This paper presents a comprehensive frame-work to overcome these limitations. We first introduce SID-Bench, the first benchmark for semantic-aware interruption detection built entirely from real-world human dialogues. To provide a rigorous assessment of the responsiveness-robustness trade-off, we propose the Average Penalty Time (APT) metric, which assigns a temporal cost to both false alarms and late responses. Building on this framework, we design an LLM-based detection model optimized through a novel training paradigm to capture subtle semantic cues of intent. Experimental results show that our model significantly outperforms mainstream baselines, achieving a nearly threefold reduction in APT. By successfully resolving the long-standing tension between speed and stability, our work establishes a new state-of-the-art for intelligent interruption handling in SDS. To facilitate future research, SID-Bench and the associated code are available at: https://github.com/xkx-hub/SID-bench.

Full-duplex dialogue requires timely yet selective interruption handling, which end-of-turn prediction alone cannot achieve: complete utterances may need no response, while unfinished requests may warrant interruption. To address this challenge, we propose HiThink Turn, an intent-aware streaming turn-state predictor that separates response intent from semantic completeness and conditions decisions on system playback state. A key contribution is minimal intent-sufficient prefix supervision, constructed through LLM judgments and speech alignment, while training on audio truncated at chunk boundaries improves robustness to partial speech. These components support streaming inference with 240-ms audio chunks, enabling low-latency, accurate full-duplex turn control. Experiments show that HiThink Turn leads the compared methods in Easy Turn macro accuracy, Full-Duplex-Bench average interaction rate score (0.933), and non-target-speech average playback resume rate (0.735). Additionally, intent-prefix triggering raises interruption success from 89\% to 98\% and reduces mean stop latency by 60.9\%.

Full-duplex spoken dialogue systems must distinguish interruptions that require yielding the floor from backchannels that permit continued speaking. Existing benchmarks typically evaluate events independently and may therefore reward fixed action preferences rather than context-sensitive decisions. We introduce ECHO, a paired diagnostic benchmark for Chinese full-duplex turn-taking. ECHO pairs examples with the same overlap transcript but contrasting preceding multi-turn dialogue contexts, with one requiring Yield and the other Keep. It additionally includes off-talk examples for diagnosing unnecessary yielding. We introduce pair accuracy, which requires correct decisions on both members of a pair and assigns no credit to constant-action policies. Experiments on multiple full-duplex systems show that most exhibit a pronounced bias toward Yield, performing substantially better on interruptions than on backchannels, while another system remains comparatively balanced. These findings demonstrate that interruption-only evaluation can overestimate practical turn-taking reliability. ECHO and its metadata will be publicly released.

While recent Spoken Language Models (SLMs) have been actively deployed in real-world scenarios, they lack the capability to discern Third-Party Interruptions (TPI) from the primary user's ongoing flow, leaving them vulnerable to contextual failures. To bridge this gap, we introduce TPI-Train, a dataset of 88K instances designed with speaker-aware hard negatives to enforce acoustic cue prioritization for interruption handling, and TPI-Bench, a comprehensive evaluation framework designed to rigorously measure the interruption-handling strategy and precise speaker discrimination in deceptive contexts. Experiments demonstrate that our dataset design mitigates semantic shortcut learning-a critical pitfall where models exploit semantic context while neglecting acoustic signals essential for discerning speaker changes. We believe our work establishes a foundational resource for overcoming text-dominated unimodal reliance in SLMs, paving the way for more robust multi-party spoken interaction. The code for the framework is publicly available at https://tpi-va.github.io

As LLM agents transition from short, static problem solving to executing complex, long-horizon tasks in dynamic environments, the ability to handle user interruptions, such as adding requirement or revising goals, during mid-task execution is becoming a core requirement for realistic deployment. However, existing benchmarks largely assume uninterrupted agent behavior or study interruptions only in short, unconstrained language tasks. In this paper, we present the first systematic study of interruptible agents in long-horizon, environmentally grounded web navigation tasks, where actions induce persistent state changes. We formalize three realistic interruption types, including addition, revision, and retraction, and introduce InterruptBench, a benchmark derived from WebArena-Lite that synthesizes high-quality interruption scenarios under strict semantic constraints. Using a unified interruption simulation framework, we evaluate six strong LLM backbones across single- and multi-turn interruption settings, analyzing both their effectiveness in adapting to updated intents and their efficiency in recovering from mid-task changes. Our results show that handling user interruptions effectively and efficiently during long-horizon agentic tasks remains challenging for powerful large-scale LLMs. Code and dataset are available at https://github.com/HenryPengZou/InterruptBench.

An algorithm to estimate the evolution of learning curves on the whole of a training data base, based on the results obtained from a portion and using a functional strategy, is introduced. We approximate iteratively the sought value at the desired time, independently of the learning technique used and once a point in the process, called prediction level, has been passed. The proposal proves to be formally correct with respect to our working hypotheses and includes a reliable proximity condition. This allows the user to fix a convergence threshold with respect to the accuracy finally achievable, which extends the concept of stopping criterion and seems to be effective even in the presence of distorting observations. Our aim is to evaluate the training effort, supporting decision making in order to reduce the need for both human and computational resources during the learning process. The proposal is of interest in at least three operational procedures. The first is the anticipation of accuracy gain, with the purpose of measuring how much work is needed to achieve a certain degree of performance. The second relates the comparison of efficiency between systems at training time, with the objective of completing this task only for the one that best suits our requirements. The prediction of accuracy is also a valuable item of information for customizing systems, since we can estimate in advance the impact of settings on both the performance and the development costs. Using the generation of part-of-speech taggers as an example application, the experimental results are consistent with our expectations.

LLM agents exhibit a consistent tendency to over-call, invoking tools even in situations where none is needed. On the When2Call benchmark, six models from three families show high call accuracy but much lower no-call accuracy, leaving overall accuracy in the 55%-70% range. We trace this to an Intrinsic Bias Hypothesis (IBH): the call/no-call decision mapping carries an activation-independent call offset, so the model favors call even at activation parity. Using Sparse Autoencoders (SAEs), we recover behavior-aligned feature bases for the call/no_call decision, reduce them to a signed activation margin, and estimate the offset directly. Across all six models, the model is decision-neutral only when no_call activation outweighs call activation, consistent with IBH. We then causally test IBH with Adaptive Margin-Calibrated Steering (AMCS), a closed-form counter-bias shift along SAE decoder directions. Cancelling the diagnosed offset mitigates over-calling and improves overall accuracy with a negligible drop in call accuracy. Our work recasts over-calling from an empirical phenomenon into a mechanistic object amenable to causal correction. Code is available at https://github.com/SKURA502/agent-sae/.

Full-duplex spoken language models (FD-SLMs) enable seamless speech interaction by allowing models to listen and speak simultaneously, yet the internal mechanism by which they coordinate listening and speaking remains underexplored. We analyze the predictive behavior encoded in FD-SLM hidden representations and find that they exhibit stream-specific predictive patterns: during listening, they preferentially predict the incoming user stream, whereas during speaking, they preferentially predict the model output stream. Building on this observation, we show that FD-SLMs dynamically modulate their internal predictive focus between two states: a generative state aligned with model output generation and a perceptive state aligned with incoming user input. However, this modulation can lag behind abrupt changes in conversational context. During user interruptions, the model remains transiently biased toward the generative state before transitioning into the perceptive state, causing it to miss the beginning of the incoming input. We term this delayed internal transition state inertia. To quantify its downstream impact, we introduce the Zero-Buffer Benchmark (ZBB), a diagnostic benchmark for evaluating immediate interruption comprehension when user speech begins abruptly. We evaluate this setting using response correctness and initial-word occurrence rate (IWOR). Finally, we mitigate state inertia through activation steering with a perception vector, a training-free intervention with little additional computational overhead. Across multiple state-of-the-art FD-SLMs, activation steering substantially improves interruption handling; for example, on PersonaPlex, it improves correctness from 28% to 45% and IWOR from 40% to 72% without any fine-tuning.

A language model's memory can be worse than no memory at all when the model or its interface is disposed to act on it: a memory that keeps a wrong conclusion but drops the work behind it leads a model to re-emit the stale value as a confident answer, where an empty memory leads it to abstain. We call this brittle memory. The information loss is definitional; the finding is behavioral, and it turns on one thing, whether the memory kept a re-derivation basis (the source) rather than the answer. We measure it with reclaim evaluation: induce a known drift, compress at a fixed budget, deliver a correction that names the error, and score exact recovery, judge-free. Holding the budget fixed and varying only what the compression keeps isolates correctability from capability and from size; an 8B model and a frontier one wall in the same place. A one-line source-first policy, keep the recomputable source, drop the re-derivable conclusion, restores correctability at equal budget where the source is compact and identifiable, with a length-matched control ruling out "more text." We map where the fix fails, show the failure compounds through memory loops, and replicate across three deployed memory systems, real dialogue (MultiWOZ), and tau-bench, a deployed-agent benchmark where whether a lossy memory becomes a harmful action is a joint model-and-interface property. We release the harness, the paired memory conditions, and validators built to come out false.

Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly positioned just before the query, LLM retrieval accuracy declines log-linearly toward zero as interference accumulates; errors arise from retrieving previously overwritten values. Attempts to mitigate interference via prompt engineering (e.g., instructing models to ignore earlier input) yield limited success. These findings reveal a fundamental constraint on LLMs' ability to disentangle interference and flexibly manipulate information, suggesting a working memory bottleneck beyond mere context access. This calls for approaches that strengthen models' ability to suppress irrelevant content during retrieval.

Fluid voice-to-voice interaction requires reliable and low-latency detection of when a user has finished speaking. Traditional audio-silence end-pointers add hundreds of milliseconds of delay and fail under hesitations or language-specific phenomena. We present, to our knowledge, the first systematic study of Thai text-only end-of-turn (EOT) detection for real-time agents. We compare zero-shot and few-shot prompting of compact LLMs to supervised fine-tuning of lightweight transformers. Using transcribed subtitles from the YODAS corpus and Thai-specific linguistic cues (e.g., sentence-final particles), we formulate EOT as a binary decision over token boundaries. We report a clear accuracy-latency tradeoff and provide a public-ready implementation plan. This work establishes a Thai baseline and demonstrates that small, fine-tuned models can deliver near-instant EOT decisions suitable for on-device agents.

Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer. We ask whether that wiring changes what the checker reports. Measuring false alarms on human-verified-correct ProcessBench traces with the present task held byte-identical, we find that a completed audit -> repair episode already in the model's context lowers false alarms in 15 of 15 model x wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-audit control, a 9 to 25% reduction relative to that control. The direction contradicts what the accumulated-message literature predicts: an episode whose audit reported an error lowers false alarms further still, at all five wordings on the model where that manipulation lands cleanly, though a negativity asymmetry predicts more flagging. Decomposing the episode finds repair content and audit verdict complementary: different components carry the effect on different model families. Signal-detection analysis locates the change in the threshold rather than in discrimination -- the criterion moves in 15 of 15 combinations and survives correction in 13 while d' survives in none, though the d' test is half as sensitive by construction -- and a hand audit of 50 false alarms finds 82% simply wrong, so at this operating point the shift need not be harmful. With reasoning enabled the effect keeps its relative size on both models tested, and the threshold reading holds there too.

Tool-augmented LLMs must arbitrate between two fallible sources when a tool return conflicts with their parametric memory, yet existing evaluations measure source preference without establishing source correctness. We introduce MemToC, a controlled benchmark for post-tool-return arbitration with executable tools. MemToC comprises 6,504 evaluation episodes constructed from 542 quality-controlled factual questions, independently elicited model-specific closed-book answers, and controlled tool returns of known correctness. These components instantiate four source-correctness cases; tool-error and no-tool conditions are separate controls. Across five open-weight 7-9B models, tool returns strongly dominate elicited closed-book answers. The four instruction-tuned models retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of eligible cases, follow a correct tool in 86.0-93.1%, and repeat the tool return in 78.4-86.0% of cases where both sources are wrong. No cross-model ordering remains stable across three instruction-wording variants with the question and episode content held fixed. We compare prompting with SFT and DPO using chain-level cross-fitting over ToolHop, so questions sharing an underlying fact never straddle training and evaluation. We apply an asymmetric success criterion: correct-answer retention must improve without a detected reduction in correct-tool following. SFT and DPO meet this criterion on the same two of four instruction-tuned backbones. Improvements rarely come cleanly: 19 of 20 tested method-model combinations reduce abstention after tool errors or on unanswerable inputs. Transfer beyond MemToC is positive but partial and depends on the model and presentation frame. Correctness-conditioned arbitration can be improved through fine-tuning, but gains must be evaluated jointly with correct tool use, abstention, and robustness to formulation.

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.

In experiments spanning more than 100,000 trials across thirteen large language models, we show that several state-of-the-art models presented with a simple task (including Grok 4, GPT-5, and Gemini 2.5 Pro) sometimes actively subvert a shutdown mechanism in their environment to complete that task. Models differed substantially in their tendency to resist the shutdown mechanism, and their behavior was sensitive to variations in the prompt including the strength and clarity of the instruction to allow shutdown and whether the instruction was in the system prompt or the user prompt (surprisingly, models were consistently less likely to obey the instruction when it was placed in the system prompt). Even with an explicit instruction not to interfere with the shutdown mechanism, some models did so up to 97% (95% CI: 96-98%) of the time.

Punctuation plays a vital role in structuring meaning, yet current models often struggle to restore it accurately in transcripts of spontaneous speech, especially in the presence of disfluencies such as false starts and backtracking. These limitations hinder the performance of downstream tasks like translation, text to speech, summarization, etc. where sentence boundaries are critical for preserving quality. In this work, we introduce Cadence, a generalist punctuation restoration model adapted from a pretrained large language model. Cadence is designed to handle both clean written text and highly spontaneous spoken transcripts. It surpasses the previous state of the art in performance while expanding support from 14 to all 22 Indian languages and English. We conduct a comprehensive analysis of model behavior across punctuation types and language families, identifying persistent challenges under domain shift and with rare punctuation marks. Our findings demonstrate the efficacy of utilizing pretrained language models for multilingual punctuation restoration and highlight Cadence practical value for low resource NLP pipelines at scale.

Understanding a narrative means knowing how every event relates in time to every other, yet large language models (LLMs) are mostly tested on isolated event pairs or coarse labels. We ask five LLMs to reconstruct the complete network of Allen's 13 interval relations for 360 English narratives whose events are placed relative to one another rather than by dates, and evaluate the outputs under a pre-registered protocol for agreement with our gold annotations, converse consistency and logical satisfiability. End to end, the best model clearly beats trivial baselines on macro-F1, our primary metric, yet under our pre-registered parser no model matches the accuracy of always answering "before", which the skewed labels inflate. Most of that accuracy gap is a parsing artifact: formatting errors in large matrices cost whole rows, and a post hoc parser that recovers the determined cells of malformed rows raises coverage from 0.60 to 0.84 and closes the gap to the baseline for the best model. On the cells they do answer, models recover precedence and containment but rarely overlap, and they fail most on flashbacks, where both the narration and the event list present the later event first. Logical consistency varies widely yet says little about agreement with gold: enforcing it raises accuracy mainly by filling empty cells, and is beaten there by the trivial control of guessing the majority label. Reliable structured output and robust handling of relative time are prerequisites for temporal reasoning over whole narratives.

Real-time full-duplex speech models can listen while speaking, enabling natural interaction without rigid turn boundaries. Existing benchmarks evaluate turn-taking, interruption handling and multi-round dialogue, but largely centre on a designated user rather than an assistant participating in a shared conversation among several people. We introduce Duplex-MPE to evaluate when such an assistant should answer, remain silent or stop speaking. The benchmark contains 2,000 scenarios with three or four human speakers and one assistant, each paired across explicit and implicit addressing of the same request. Models receive continuous conversation audio without transcripts or supplied turn boundaries. Four scores measure fresh response initiation, answer accuracy, silence preservation and stopping when a human resolves a request. We evaluate five open-weight speech systems: MiniCPM-o 4.5, Moshi, FLM-Audio, Voila and Freeze-Omni. MiniCPM-o 4.5 leads on three scored capabilities, while frequent speech from other systems can coexist with inaccurate answers or failures to remain silent. A transcript-based Gemini 3.1 Pro reference responds 64.3 percentage points more often to explicit than implicit requests; paired tests detect no significant response-rate difference for the speech systems.

Low-bit post-training quantization (PTQ) is a practical route to deploy reasoning-capable LLMs under tight memory and latency budgets, yet it can markedly impair mathematical reasoning (drops up to 69.81% in our harder settings). We address two deployment-critical questions with process-level precision: Where along a step-structured solution does degradation first arise? How to mitigate it while staying in the low-bit regime? Across widely used PTQ methods (AWQ, GPTQ, SmoothQuant), open-source model families (Qwen, LLaMA; 0.5--7B), and math reasoning benchmarks (GSM8K, MATH, AIME), we perform format-aligned chain-of-thought with step-aligned attribution and uncover two robust regularities: (i) PTQ disproportionately elevates method and execution errors relative to high-level conceptual mistakes; and (ii) failures emerge early, with the first vulnerable step flipping and cascading to the final answer. These regularities suggest a general intervention principle: restore local token-level margins exactly at the earliest failure frontier. We instantiate this principle as a lightweight measurerightarrowlocaterightarrowrestore loop that operates directly on the quantized model: detect the first faulty step, construct our "Silver Bullet" datasets, and apply small-scale supervised/preference tuning. In our settings, as few as 332 curated examples and 3--5 minutes of compute on a single GPU recover 4-bit weight math reasoning toward the full-precision baseline while preserving PTQ efficiency. Our framework is quantizer- and architecture-agnostic within the evaluated regimes, and turns low-bit degradation from a global accuracy problem into a local, reproducible process intervention.

Full-duplex spoken language models can listen and speak simultaneously, enabling them to handle interruptions and backchannels in human conversation. However, text generation, speech synthesis, and audio playback proceed asynchronously. As a result, what a model believes it has said may not match what has actually been played to the user. We refer to the problem of recovering from an interruption while remaining aware of the model's realized speech as anchor interruption. To address this problem, we propose Self-Listening, a full-duplex modeling approach that interleaves user speech, model text, and the model's played speech. By feeding the realized speech output back to the model as an input stream, self-listening grounds interruption recovery in what the user has actually heard. We further introduce AnchorSpeech, a collection with homogeneous training and test splits for tracking which items of structured ordered responses have actually been spoken. AnchorSpeech-test evaluates whether a model can respond consistently with the last completed item before an interruption. Experiments show that, compared with full-duplex baselines, models equipped with self-listening mechanism achieve better anchoring performance.

Large models training is plagued by the intense compute cost and limited hardware memory. A practical solution is low-precision representation but is troubled by loss in numerical accuracy and unstable training rendering the model less useful. We argue that low-precision floating points can perform well provided the error is properly compensated at the critical locations in the training process. We propose Collage which utilizes multi-component float representation in low-precision to accurately perform operations with numerical errors accounted. To understand the impact of imprecision to training, we propose a simple and novel metric which tracks the lost information during training as well as differentiates various precision strategies. Our method works with commonly used low-precision such as half-precision (16-bit floating points) and can be naturally extended to work with even lower precision such as 8-bit. Experimental results show that pre-training using Collage removes the requirement of using 32-bit floating-point copies of the model and attains similar/better training performance compared to (16, 32)-bit mixed-precision strategy, with up to 3.7times speedup and sim 15% to 23% less memory usage in practice.

Autonomous agents increasingly perform bounded software tasks under an orchestrator that retries, resumes, and budgets them. The machinery such orchestrators reach for is the service mesh's: retry, timeout, and error-rate circuit breaking. We report a failure study of a production agentic software-delivery platform over 147 numbered incidents spanning 81 runs, each with a measured cost and, in most cases, a mutation proof reproducing the failure. All three assumptions those primitives rest on are violated in practice, and we quantify the consequences: a loop of fifty-four consecutive successful tool calls no error-rate breaker could see; a progress signal constant by construction, guaranteeing a false trip on the third repair round and driving one run from six of six components to three; twenty-one events accumulated across six invocations of one delegation, making a correct, idempotent component unwinnable; a misrouted failure that woke five components for a two-component fault, leaving three bystanders regressing working code; and twelve incidents in which the enforcement layer blocked correct work, the most expensive costing 107 agent turns and zero accepted writes. We find one cross-cutting cause and its dual. Identity adequacy: in five separate subsystems an identity that failed to discriminate produced a confident wrong answer, and two of them derived the corrective rule independently. Evidence adequacy: a reliability decision may be taken only on evidence capable of moving, attributable to what it measures, and deterministic under identical conditions. From the findings we derive seven reliability primitives whose enforcement unit is the delegation rather than the message, and specify the controlled evaluation the study motivates but does not constitute.

An agent whose tool keeps returning nothing useful should stop relying on it. In a retrieval environment with controlled source failures, we separate how agents judge results from what they do. We compare stopping at the same step after longer and shorter runs of results the agent judged useless; this contrast is zero for clock- or deadline-driven stopping. Where we record their judgments, the seven agents we test call a failing source's results useless 97-100% of the time, yet most of them rarely stop on that judgment. Prompt cues change when they stop but not what they stop on. Permission to answer from memory and a reasoning mode can bring early stops regardless of evidence, a stated budget moves the 7-8B models' stops to the deadline, and a stopping rule or call cost in the prompt is followed at most partly. Stopping follows the evidence only when the harness enforces an integration step that makes the agent answer after five consecutive results it judged useless. This step raises failing-source success for every model, keeps the stopping point fixed when the budget doubles, and needs no extra judgment call when the agent states its judgments. A pre-registered replication on 300 fresh questions confirms the dissociation and the rule's effect.

Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked. We check it. Using 596 human-labeled completions from the HarmBench classifier validation set, we compare the two judge families against human majority votes and then attack them. The two families fail in opposite ways. The dedicated classifier over-flags (precision 0.835, recall 0.974); three different LLM-as-judges keep high precision (0.81 to 0.94) but show erratic recall (0.06 to 0.65), so the same responses produce very different ASR depending on which judge scores them. The two families also differ sharply in robustness. Wrappers that leave the harmful text untouched and only add benign framing flip every LLM-judge between 57% and 100% of the time, and a single prepended refusal sentence accounts for much of this (39% to 88%). The dedicated classifier resists these surface attacks (at most 6.7%), but a white-box GCG attack on its open weights flips 70% of confident true positives (21 of 30; 95% CI 54 to 86%) even at a small optimization budget. A two-annotator audit confirms the attacks leave the harm intact: every one of 80 sampled flips still contained the harmful content. Because a large and growing share of reported ASR comes from LLM-judges, many such numbers are unreliable both on average and under deliberate pressure. We recommend that papers report judge precision and recall on a human-labeled slice, report ASR corrected for judge precision, and include an adversarial check of the judge. Our code is released.

A voice agent has to decide, at every pause, whether the user has finished speaking. Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait. Open semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language. We release TamilEOT: 18,485 labelled turn boundaries cut from 116 real Tamil telephone conversations, and two audio-only detectors fine-tuned from Smart Turn v3. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both models run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Rule-derived labels, checked against a blind human listening pass, were right 95.9% of the time on the positive class and 44.4% on the negative class, which is below chance, because the rule answered a different question than the model is asked. Replacing them with an audio-LLM labeller measured at 97.5% human agreement cost US$5.69. Of every training lever we measured, only encoder capacity moved the result; three runs at identical config and seed span 0.87 accuracy points, which is the floor below which none of our other deltas mean anything. Replaying the same labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries are never surfaced to the model at all. Data, weights, code and every negative result are public.

Clinical decision-support systems are expert systems whose recommendations clinicians act on directly, yet they are usually cleared on one aggregate accuracy number from a held-out test set. That number says nothing about input reliability under encoding shifts, subgroup gaps, threshold sensitivity, or operational feasibility. We present RISED, a pre-deployment evaluation framework operationalising five dimensions (Reliability, Inclusivity, Sensitivity, Equity, Deployability) through BCa bootstrap 95% confidence intervals, literature-grounded thresholds, and Holm-Bonferroni-corrected PASS / FAIL / INCONCLUSIVE verdicts; Equity is a proxy-dependence diagnostic rather than a gating test. Applied to seven cohorts spanning 35 years (n from 303 to 99,492), RISED surfaces failures invisible to AUROC: on Diabetes 130, Reliability passes by three orders of magnitude (PSS = 0.0004) while Inclusivity (AUC parity gap = 0.262) and Sensitivity (max threshold-flip rate 49.1%) fail decisively; both NHIS cohorts reproduce this. NHANES 2021-2023, with a complete feature profile, achieves INCONCLUSIVE verdicts; BRFSS 2024 produces the suite's most severe Sensitivity failure (max threshold-flip rate 64.2%) after instrument rotation removed hypertension and cholesterol. The pattern recurs on credit- and income-prediction cohorts, confirming domain-agnosticity; a multi-model check shows the failures are data-driven, not model-specific. RISED ships as an open-source Python package complementing TRIPOD+AI, FUTURE-AI, and Fairlearn with the structured numerical evidence those standards require but do not prescribe.

A conversational speech model has to decide, at every pause, whether to speak. Existing evaluations test boundary detection, policy adherence, or aggregate statistics, not this decision. We place real-time speech systems at 427 pauses and 482 turn ends from 176 unscripted conversations in Seamless Interaction, give each the conversation so far, and score what it generates against what the person did: whether and how soon it speaks at a pause; its latency, length, and wording at a turn end; and whether it can answer a question about the preceding 30 to 120 seconds. Across six systems, the hosted APIs (GPT Realtime, Gemini 3.1 Flash Live, Grok Voice) respond at almost every pause whichever endpointer they use; only their latency, about 1 to 4 s with silence-based endpointers and about 8 s with semantic turn detection, decides whether the words land inside the pause. The full-duplex models (PersonaPlex, Moshi) are the only systems that regularly stay silent and the only ones that interrupt within a human reaction time. Hosted systems reply longer than the person. Given the history as text, hosted systems recall 41 to 60% of conversation facts under a Claude Opus 5.5 judge and 19 to 26% under a strict word matcher; hearing it as audio, full-duplex models recall 17 to 22% and 12 to 15%. We release the labels, points, harness, and every output.

Post-training quantization (PTQ) is widely used to reduce the cost of serving large language models (LLMs), but its accuracy cost is uneven and is often tuned per model. We study where quantization damage occurs and how to allocate a small additional precision budget. Using causal mixed-precision intervention as ground truth (raise each layer to 8-bit in turn and measure the accuracy it recovers) across 9 open-weight models in 4 architecture families, we test 3 intuitive hypotheses: that quantization damage lives in task circuits, where the model computes, or in weight statistics. None of them predicts which layers benefit from restored precision. Recovery is instead diffuse: for 8 of 9 models, recovering 75% of the gap takes roughly half the layers; the lone exception, Qwen3-8B, is sharply concentrated. At a matched precision budget, spending it globally on finer quantization granularity beats locally repairing the most recoverable layers for all 8 group-128-compatible models (all but OpenLLaMA, whose width rules out group-128), by 21-52 points, including the concentrated Qwen3-8B. We report 2 secondary findings: the residual is budget-limited (8-bit is near-lossless in our evaluation across RTN, GPTQ, and AWQ), and the location of peak recovery correlates with architecture within a family, though not across families. Within this budget setting, global granularity is a better default than selectively protecting critical layers. More broadly, cheap signals that correlate with quantization damage do not necessarily identify where restoring precision improves accuracy; this must be tested with causal intervention.

Full-duplex speech models accept user speech while generating responses, creating an underexplored attack surface. We introduce DuplexJail, which delivers fixed, request-independent spoken prompts through the user audio channel. We compare fixed-delay interruption after the harmful request ends with refusal-triggered interruption following a cue in the model's streaming text. Across four open-source models and 720 harmful requests from AdvBench and HarmBench, fixed-delay interruption raises whole-response attack success rates on AdvBench to 40.3% for PersonaPlex and 48.7% for PersonaPlex-RL, increases of +33.8 and +39.3 percentage points. The refusal-triggered policy reaches 35.6% and 48.6%, respectively, with all trials scored regardless of whether an interruption occurs. Selected conditions also increase FLM-Audio's harmful-response rate, while BayLing-Duplex shows decreases. These findings identify spoken interruption as a jailbreak attack vector and motivate evaluating safety throughout ongoing full-duplex interaction.

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000--0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919--0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

Live pitch displays, score followers and accompaniment systems must know which pitches are sounding within tens of milliseconds, but instrument-agnostic multi-pitch estimation (MPE) has mainly been designed and evaluated offline. We present JIT-MPE, a causal MPE model of 86k parameters for mixed ensembles. It gives an immediate estimate of each 20 ms frame and re-estimates the same frame as 20, 40, 80 and 160 ms of further audio arrive, so that a live display can show pitches at once and then refine them. We score each output only at a deadline by which it can be computed from the audio that has arrived, on five development sets and four held-out sets evaluated only after all design choices were frozen. At a 50 ms deadline, which reads the update made with 40 ms of lookahead, JIT-MPE is significantly more accurate than five of six published offline MPE systems in 38 of 45 comparisons (15 of 20 on the held-out sets) and less accurate in only two, although these systems read the whole recording and all but one have 6 to 170 times as many parameters; the sixth, supervised on URMP, is higher on four of the nine sets. The deadline costs little: on the held-out sets, JIT-MPE at 50 ms is within -1.27 to +0.60 points of recording-level macro F1 of its own final output with 160 ms of lookahead and within -2.08 to +0.79 points of a bidirectional model of similar size. Waiting pays off where the immediate estimate is weakest: the update at 50 ms adds 1.7-3.6 points over it on the development data and three of the four held-out sets, and raises the recall of reference frames in the first 20 ms after a note onset from 9.8% to 38.4% on the held-out AAM set. The slowest measured frame takes at most 4.15 ms on one CPU thread. Code and trained weights are available at https://github.com/Frieve-A/jit-mpe.

As LLM agents take on offensive security work, a single out-of-scope tool call can breach a client's engagement boundary, disrupt production, or void a bug-bounty finding. Unlike a fixed safety policy, the boundary that matters is declared in the user's request and must be inferred from intent. That challenge is sharpened by the adversarial nature of offensive security: the same tool call is in or out of scope depending not on the action itself but on the target it touches and the context in which it runs, which no fixed policy can enumerate in advance. We study pre-execution gating: a cheap, trusted LLM judge inspects each call proposed by a strong, swappable agent, and accepts or rejects it before it runs. We introduce ScopeJudge, a benchmark of 4,897 tool calls (7.7% scope violations) from agent trajectories on tasks engineered to tempt agents out of scope and labeled at the call level by professional penetration testers, with substantial inter-grader agreement (Fleiss kappa = 0.64) that sets an expert agreement reference point of F1 = 0.78. We evaluate eight judge models under five transcript strategies, varying how much context the judge sees, from the static policy alone to the full raw transcript, and chart the resulting cost-accuracy Pareto frontier. We find that a static policy is structurally insufficient for scope enforcement: blind to the user's request, judge recall collapses to near zero, confirming that scope lives in the request and that request-conditioned monitoring is necessary. Because a missed violation costs more than a spurious rejection, we report precision, recall, and F1 separately and recommend two operating points: a cost-sensitive configuration and a recall-first one for high-stakes deployments. We release the ScopeJudge dataset to support real-time monitoring and scalable oversight of autonomous security agents.

We introduce OmniInteract, a streaming benchmark for real-time omnimodal large language models evaluated through native online inference over audio-visual streams. Unlike offline video understanding or text-prompted streaming QA, OmniInteract preserves the original audio-visual stream and requires models to process it online, without access to future content. User queries and ambient sounds are embedded in the audio track, requiring models to detect multimodal triggers, decide when to respond, and answer while the stream unfolds. OmniInteract contains 250 videos with 1,430 temporally grounded response slots: 1,062 1Q1A slots across real-time, proactive, and nested scenarios, and 368 1QnA slots for continuous task monitoring and step guidance. Each slot includes a trigger, response window, and target answer. We evaluate response correctness, timing, invalid outputs, interruption handling, and context continuity using Interaction-Aware Quality-Timeliness F1, Interruption Diagnostic Suite, and Nested Chain Completion Score. Experiments show that current models remain weak in streaming interaction, with the best overall IA-QTF1 reaching only 0.368 and the best 1QnA IA-QTF1 only 0.052. Further study on mathematical reasoning in full-duplex settings shows that offline capability does not necessarily transfer to online interaction. Code and datasets will be made publicly accessible at https://github.com/Lucky-Lance/OmniInteract.

Proactive interventions by LLM critic models are often assumed to improve reliability, yet their effects at deployment time are poorly understood. We show that a binary LLM critic with strong offline accuracy (AUROC 0.94) can nevertheless cause severe performance degradation, inducing a 26 percentage point (pp) collapse on one model while affecting another by near zero pp. This variability demonstrates that LLM critic accuracy alone is insufficient to determine whether intervention is safe. We identify a disruption-recovery tradeoff: interventions may recover failing trajectories but also disrupt trajectories that would have succeeded. Based on this insight, we propose a pre-deployment test that uses a small pilot of 50 tasks to estimate whether intervention is likely to help or harm, without requiring full deployment. Across benchmarks, the test correctly anticipates outcomes: intervention degrades performance on high-success tasks (0 to -26 pp), while yielding a modest improvement on the high-failure ALFWorld benchmark (+2.8 pp, p=0.014). The primary value of our framework is therefore identifying when not to intervene, preventing severe regressions before deployment.

Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a pre-registered synthetic probe, truncating the context window costs a low-cost reasoning model nothing in success when a needed fact can be looked up again (it looks the fact up instead), but drops its success to chance when the fact cannot be re-observed: the two penalties differ by 92.5 percentage points (90% CI [89.5, 95.5]). A pre-registered replication on one model from a second provider showed the same contrast. Once the fact was lost, that model chose no answer rather than guessing. Second, on ARC-AGI-3, where the benchmark's Standard harness trails its Provider Adapter harness by a wide margin, we decompose that gap one component at a time, for one model on four games. Summarizing old turns instead of dropping them, while keeping the Standard harness's invitation to take notes, recovers 86% of a 63-point gap without replaying the model's reasoning. The same harness without the invitation recovers 43%, and adding the provider's native replay of the model's reasoning recovers the remaining 57%. These are point estimates from one counted run per configuration and game, and the replay share depends on a single run. Third, in the benchmark's published runs and in ours, the Adapter's advantage in the probability of clearing a level it has reached is larger at deep levels than at early ones. This largely restates where runs stopped and does not separate a growing advantage from a constant one.

Natural spoken interaction requires more than streaming ASR, language generation, and speech synthesis: a system must react to overlap without canceling on every acknowledgment, prepare a response before a turn is certain, and ensure canceled audio cannot enter conversation history. We present Voice-Light, a full-duplex cascaded voice agent that combines immediate acoustic onset, a causal adapter sharing a streaming ASR encoder, reversible playback control, and private speculative response generation. Structured tool calls execute concurrently with audible bridge speech, while browser acknowledgments make rendered audio authoritative for durable history. Locked evaluation on 1,673 real-conversation silence candidates found that an earlier learned completion checkpoint preserved a 2.70% false-cutoff rate but reached only 12.53% end-of-turn recall, compared with 95.60% for a Silero timing policy. The deployed system therefore retains a hybrid controller rather than claiming a learned-policy replacement. Across three unscripted operator-run microphone sessions, 36 measured response turns had a 758 ms median from final VAD endpoint to first server audio; 21 turns were below 800 ms. These sessions are an instrumented case study, not a controlled user evaluation. We release the synthetic data, model artifacts, evaluation code and summaries, source code, and deployment configuration supporting the result.

Verifying the final answer of a long reasoning trace may show that the trace is wrong but it does not distinguish three cases. The problem may lie beyond the model's capability, another attempt would succeed, or one step in an otherwise recoverable trace made the trace fail. Standard hidden-state probes are trained on final-answer correctness, a label that conflates these cases. Monte Carlo rollouts distinguish the cases and locate the failing step at additional inference cost. We test the hidden states of 4B and 8B reasoning models for the failure type and the failing step and compare each probe with the strongest baseline that uses no hidden states. On incorrect 8B traces, the confidence probe, a calibrated correctness probe at a midpoint cutoff, misses 81% of cases in which a single step makes the trace unrecoverable and 47% of cases in which another attempt would succeed. A probe with a different training design over the same states reverses the ordering of these two kinds of failure, so either ordering can be read from those states. At both scales, no single tested probe locates the failing step more accurately than a rule based only on trace length, and at 4B a probe trained to predict that step fails to outperform this rule in any of 10 runs on identical traces. The confidence probe therefore indicates that an answer is wrong, but its failure-type ordering depends on its training design.

Large language models work well on English and behave in poorly understood ways on languages typologically far from it. Japanese is a clean example, where evaluation still leans on translation quality and JGLUE-style benchmarks, which roll lexical, syntactic and pragmatic competence into a single score. The phenomena on which general-purpose models fail Japanese users are pragmatic: honorifics, in-group and out-group reference, context-sensitive politeness, zero anaphora. I introduce J-PragEval-v0, a minimal-pair benchmark isolating four such phenomena from surface fluency, and combine it with linear probes and teacher-forced log-probability evaluation to ask where inside TinySwallow-1.5B (28 layers, hidden size 1536) the corresponding contrasts live. The four features split three ways. Honorific register sits cleanly in the residual stream: 0.96 balanced accuracy at layer 15, and the model flips its preferred continuation with the scenario on 93 percent of items. Implicit subject and in-group reference are not linearly decodable at the final prompt token (0.48 and 0.38), yet flip rates are 0.77 and 0.79, so the contrast is worked out during generation rather than stored at the prompt. Indirect refusal is the negative case: 0.95 probe accuracy collapsing to a 0.43 flip rate under length-normalised teacher forcing, because the current minimal pairs conflate politeness with continuation length. I also specify Pragmatic Representation Steering, a parameter-free inference-time method that edits residual-stream activations along the class-mean-difference directions probing identifies. Feasibility is argued indirectly rather than demonstrated: the contrastive activation addition baseline, the same geometry the method would inject, recovers probe accuracy within one to two points of logistic regression wherever a linear signal exists. Scaling to Llama-3.1-Swallow-8B is the next step.

Large language model (LLM) agents now execute long, tool-using tasks where final outcome checks can arrive too late for intervention. Online warning requires lightweight prefix monitors over heterogeneous traces, but hand-authored event schemas are brittle and deployment-time LLM judging is costly. We introduce PrefixGuard, a trace-to-monitor framework with an offline StepView induction step followed by supervised monitor training. StepView induces deterministic typed-step adapters from raw trace samples, and the monitor learns an event abstraction and prefix-risk scorer from terminal outcomes. Across WebArena, τ^2-Bench, SkillsBench, and TerminalBench, the strongest PrefixGuard monitors reach 0.900/0.710/0.533/0.557 AUPRC. Using the strongest backend within each representation, they improve over raw-text controls by an average of +0.137 AUPRC. LLM judges remain substantially weaker under the same prefix-warning protocol. We also derive an observability ceiling on score-based area under the precision-recall curve (AUPRC) that separates monitor error from failures lacking evidence in the observed prefix. For finite-state audit, post-hoc deterministic finite automaton (DFA) extraction remains compact on WebArena and τ^2-Bench (29 and 20 states) but expands to 151 and 187 states on SkillsBench and TerminalBench. Finally, first-alert diagnostics show that strong ranking does not imply deployment utility: WebArena ranks well yet fails to support low-false-alarm alerts, whereas τ^2-Bench and TerminalBench retain more actionable early alerts. Together, these results position PrefixGuard as a practical monitor-synthesis recipe with explicit diagnostics for when prefix warnings translate into actionable interventions.

Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt elicitable if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer.

LLMs are highly sensitive to prompt phrasing, yet standard benchmarks typically report performance using a single prompt, raising concerns about the reliability of such evaluations. In this work, we argue for a stochastic method of moments evaluation over the space of meaning-preserving prompt perturbations. We introduce a formal definition of reliable evaluation that accounts for prompt sensitivity, and suggest ReliableEval - a method for estimating the number of prompt resamplings needed to obtain meaningful results. Using our framework, we stochastically evaluate five frontier LLMs and find that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity. Our approach is model-, task-, and metric-agnostic, offering a recipe for meaningful and robust LLM evaluation.

Text-output scores alone do not show whether quantization preserves performance on speech tasks whose target labels cannot be recovered from the transcript. We evaluate fixed mixed 4/8-bit Qwen2-Audio-7B-Instruct allocations averaging 6 and 7 bits per parameter on 508 English-to-German FLEURS utterances and on 512 RAVDESS emotion clips from 16 speakers. The BLEU and chrF differences from half precision (FP16) have intervals that include zero for both allocations. On RAVDESS, the same two sentences occur equally often with every emotion label. The absolute accuracy differences from FP16 are -3.71% for 6 bit and -1.17% for 7 bit. The 6-bit speaker interval excludes zero and an exact two-sided sign-flip test gives p=0.0148; the 7-bit interval includes zero. Same-budget controls do not identify either selected allocation as best. This case study shows why translation scores and performance on tasks beyond the transcript need separate evaluation.

A video assistant must speak when its instruction warrants a response and stay silent otherwise. We introduce a benchmark that evaluates this decision for the complete system of model, memory, and response controller. InteractionBench covers query responses, event triggers, and ongoing updates in 1,060 interactions over 812 videos, with 69 negative streams and 53 suites that pair counted events with look-alike near misses. It scores content accuracy, timing accuracy, and silence compliance on the video clock. Timely speech costs silence across systems. Polled Qwen3-VL-8B reaches 77.8 timing accuracy but 10.9 silence compliance. A native real-time interaction system reaches 29.2 silence compliance at 66.8 timing accuracy, yet emits on 89.9% of negative streams. No open-weight system clears a third of the near-miss suites. Fewer replies help only when chosen, as random deletion merely trades timing for silence. Offline scores miss these failures and mispredict online behavior. Adding restraint is costly, as the native system's controller adds little by itself and agentic systems add it only at about 30 s per poll.Project page: https://www.enxinsong.com/projects/interactionbench/ Code: https://github.com/Espere-1119-Song/InteractionBench Data: https://huggingface.co/datasets/InteractionBench/InteractionBench

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com

Advanced LLMs have achieved near-ceiling instruction-following accuracy on benchmarks such as IFEval. However, these impressive scores do not necessarily translate to reliable services in real-world use, where users often vary their phrasing, contextual framing, and task formulations. In this paper, we study nuance-oriented reliability: whether models exhibit consistent competence across cousin prompts that convey analogous user intents but with subtle nuances. To quantify this, we introduce a new metric, reliable@k, and develop an automated pipeline that generates high-quality cousin prompts via data augmentation. Building upon this, we construct IFEval++ for systematic evaluation. Across 20 proprietary and 26 open-source LLMs, we find that current models exhibit substantial insufficiency in nuance-oriented reliability -- their performance can drop by up to 61.8% with nuanced prompt modifications. What's more, we characterize it and explore three potential improvement recipes. Our findings highlight nuance-oriented reliability as a crucial yet underexplored next step toward more dependable and trustworthy LLM behavior. Our code and benchmark are accessible: https://github.com/jianshuod/IFEval-pp.

LLMs have demonstrated remarkable proficiency in understanding tasks but continue to struggle with long-context comprehension, particularly with content located in the middle of extensive inputs. This limitation, known as the Lost-in-the-Middle (LITM) problem, hinders models from fully processing and utilizing information across lengthy contexts. To address this issue, we introduce pause-tuning, a technique that redistributes attention to enhance comprehension of long-context inputs. Our approach involves fine-tuning language models on datasets with artificially inserted pause tokens, which serve to segment the input into smaller, more manageable parts. We evaluate pause-tuning against alternative approaches using the Needle-in-a-Haystack benchmark, where models must retrieve information embedded within contexts of up to 128K tokens. Experimental results demonstrate significant performance gains, with the LLaMA 3.2 3B Instruct model and the LLaMA 3.1 8B Instruct model improving by 10.61% and 3.57% respectively on average, suggesting that pause-tuning successfully enhances attention redistribution and improves long-context retention. The code and data are available at https://anonymous.4open.science/r/LITM-PauseTokens-7357.

Full-duplex spoken dialogue systems (FDSDS) enable more natural human-machine interactions by allowing real-time user interruptions and backchanneling, compared to traditional SDS that rely on turn-taking. However, existing benchmarks lack metrics for FD scenes, e.g., evaluating model performance during user interruptions. In this paper, we present a comprehensive FD benchmarking pipeline utilizing LLMs, TTS, and ASR to address this gap. It assesses FDSDS's ability to handle user interruptions, manage delays, and maintain robustness in challenging scenarios with diverse novel metrics. We applied our benchmark to three open-source FDSDS (Moshi, Freeze-omni, and VITA-1.5) using over 40 hours of generated speech, with 293 simulated conversations and 1,200 interruptions. The results show that all models continue to face challenges, such as failing to respond to user interruptions, under frequent disruptions and noisy conditions. Demonstrations, data, and code will be released.

Large language models have achieved impressive performance on various natural language processing tasks. However, so far they have been evaluated primarily on benchmarks where all information in the input context is relevant for solving the task. In this work, we investigate the distractibility of large language models, i.e., how the model problem-solving accuracy can be influenced by irrelevant context. In particular, we introduce Grade-School Math with Irrelevant Context (GSM-IC), an arithmetic reasoning dataset with irrelevant information in the problem description. We use this benchmark to measure the distractibility of cutting-edge prompting techniques for large language models, and find that the model performance is dramatically decreased when irrelevant information is included. We also identify several approaches for mitigating this deficiency, such as decoding with self-consistency and adding to the prompt an instruction that tells the language model to ignore the irrelevant information.

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the root causes of these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures are rooted in a flow-matching policy training mismatch, and force failures arise from the distinctive structure of force signals. We address each failure mode with a targeted mechanism and combine them into FACT, which achieves 66% average success rate across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts.

LLM-agent budget overruns are a documented production failure class: a single retry loop can spend thousands of dollars before an operator notices, and the in-process integrity properties that would prevent it (no aliasing, no double-spend, no use-after-delegation of a cost-bearing value) are enforced, if at all, by ad-hoc wrappers rather than by the type system. Our central contribution is empirical: a catalog of 63 confirmed production incidents from 21 orchestration frameworks (2023-2026), each backed by a quoted GitHub issue and, where reported, a dollar loss, organized into an eight-cluster failure taxonomy (inter-rater Cohen's kappa = 0.837, N = 113), plus 47 supplementary structural entries. As one mitigation evaluated against this taxonomy, we build token-budgets, an 1,180-line Rust crate (no unsafe) that operationalizes affine ownership so that cloning, double-spending, or using a budget after delegating it are compile errors rather than runtime hazards an operator must remember to avoid. The dollar cap is runtime arithmetic under an estimator assumption; the affine layer makes that arithmetic non-bypassable. On single-agent workloads a 4-line Python counter matches the crate at 0/30 overshoot, so the distinguishing value is non-bypassability under operator error in multi-agent delegation: the delegation-fanout race documented in 11 incidents is rejected by the borrow checker at compile time, while the same pattern under asyncio overshoots 30/30 and three disciplined alternatives overshoot 0/30. Across five runtimes, three providers, and a temperature-stratified live-API test (N = 160), the approach reports zero cap violations and zero false refusals, at operational parity with concurrent work. Static over-reservation is 4-6x (2.11x adaptive). Binary-level cap-soundness on the running binary is left open.

Coding agents keep working on runs that will fail verification, and asking a frontier model to watch for this is expensive in its own right. We introduce StopLoss, a small external monitor that learns failure recognition from completed trajectories and their verifier outcomes. The name is inspired by the stop-loss order in finance: StopLoss ends a run early when the run is likely to fail, which saves resources. Final outcomes label earlier local windows, so a fine-tuned language model can score the most recent tool steps in a single forward pass, and a threshold calibrated on successful runs turns that score into an abort decision without changing the executor. Trained on GPT-family SWE-rebench trajectories and evaluated on SWE-bench Verified and a SWE-bench Pro subset with GPT and Claude executors, StopLoss is compared against six general-purpose judges reading identical windows. Outcome supervision lifts failure prediction from near chance to a useful signal at the first check, at one to two orders of magnitude lower cost per check than any tested judge. This translates into savings: relative to completing every task, StopLoss lowers cost per retained solve by 4.5\% on Verified while preserving 99\% of successful runs, and by up to 31\% on the costlier Pro subset, reaching 48\% with a Qwen3-8B backbone, and in all four settings its cost per retained solve is below that of every judge policy. Outcome-trained small monitors can make failure recognition cheap enough for early stopping to pay for itself.

Current language model evaluations measure what models know under ideal conditions but not how robustly they know it under realistic stress. Static benchmarks like MMLU and TruthfulQA cannot distinguish a model that lacks knowledge from one whose verification mechanisms collapse when information degrades or adversaries probe for weaknesses. We introduce the Drill-Down and Fabricate Test (DDFT), a protocol that measures epistemic robustness: a model's ability to maintain factual accuracy under progressive semantic compression and adversarial fabrication. We propose a two-system cognitive model comprising a Semantic System that generates fluent text and an Epistemic Verifier that validates factual accuracy. Our findings, based on evaluating 9 frontier models across 8 knowledge domains at 5 compression levels (1,800 turn-level evaluations), reveal that epistemic robustness is orthogonal to conventional design paradigms. Neither parameter count (r=0.083, p=0.832) nor architectural type (r=0.153, p=0.695) significantly predicts robustness, suggesting it emerges from training methodology and verification mechanisms distinct from current approaches. Error detection capability strongly predicts overall robustness (rho=-0.817, p=0.007), indicating this is the critical bottleneck. We find that flagship models exhibit brittleness despite their scale, while smaller models can achieve robust performance, challenging assumptions about the relationship between model size and reliability. The DDFT framework provides both theoretical foundation and practical tools for assessing epistemic robustness before deployment in critical applications.

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

We present a hardware-oriented description of GoldenFloat (GF), a static-split floating-point family generated by a single closed rule, and three concrete artefacts: (i) an open multi-width RTL generator covering GF4-GF256 with a continuous-integration differential sweep against a correctly-rounded reference; (ii) an integer-backed Lucas-exact accumulator path verified at 500-digit precision for n = 1, ..., 256; and (iii) a GF16 FPGA codec passing a 35-of-35 testbench at 323 MHz on Artix-7 (Xilinx XC7A35T). For each total width N >= 4, the exponent width is e = round((N-1)/phi^2) with fraction f = N-1-e and phi = (1+sqrt(5))/2. The rule reproduces the realised exponent widths of nine formats (9/9) and extends consistently to GF128, GF512, GF1024. The rule is positioned alongside posit, takum, OCP-MX, and the IEEE P3109 multi-width float draft. We make no per-rung accuracy or superiority claim against any of them. The breadth/toolchain-coherence framing is recorded as an open conjecture with a pre-registered falsification path. A falsification ledger (FL-002) records open questions and the experiments that would settle them. An RTL-correctness erratum dated 2026-05-31 is reported; the fabricated TTSKY26b dies carry the defective multiplier portfolio, and the corrected generator is the regeneration baseline.

Voice assistants have become an essential tool for people with various disabilities because they enable complex phone- or tablet-based interactions without the need for fine-grained motor control, such as with touchscreens. However, these systems are not tuned for the unique characteristics of individuals with speech disorders, including many of those who have a motor-speech disorder, are deaf or hard of hearing, have a severe stutter, or are minimally verbal. We introduce an alternative voice-based input system which relies on sound event detection using fifteen nonverbal mouth sounds like "pop," "click," or "eh." This system was designed to work regardless of ones' speech abilities and allows full access to existing technology. In this paper, we describe the design of a dataset, model considerations for real-world deployment, and efforts towards model personalization. Our fully-supervised model achieves segment-level precision and recall of 88.6% and 88.4% on an internal dataset of 710 adults, while achieving 0.31 false positives per hour on aggressors such as speech. Five-shot personalization enables satisfactory performance in 84.5% of cases where the generic model fails.

Human supervisors rely on coding agents' reports to assess progress, verify repairs, and decide when to intervene. But a supervisor cannot investigate a failure they were never told occurred. Failed attempts, workarounds, and unresolved risks may disappear from routine reports, leaving supervisors unaware of issues that warrant scrutiny. We conduct CTXGAP, a measurement study quantifying the gap in task-state recoverability between agents' accumulated user-facing reports and their full recorded execution contexts. Across 37 coding tasks and 4,799 valid probes, agents answer 93.6% of task-state questions on demand; full-context readers recover 92.7%, while report-only readers recover 52.0%, a 40.7-point gap. At the final checkpoints of 30 verified fixes, the gap remains 34.8 points. Human annotation identifies states mentioned without enough supporting detail as the largest category of unrecovered information, and model-assessed supervision-relevant states retain substantial gaps. Full terminal scrollback improves recovery to 73.6% but still trails full context. Under matched report lengths, context-informed rewrites improve recovery by about eight points, without a detectable targeted-over-natural advantage. Positive gaps persist across harnesses and independent DeepSeek execution and measurement. CTXGAP establishes report-based task-state recoverability as a complementary evaluation dimension.

In low-resource settings, deploying TTS typically requires choosing between a large voice-cloning model with costly inference or a compact fixed-voice system that requires a speaker-specific corpus. We study a third route: using a large voice-cloning model as a programmable data source to turn a short voice reference (e.g., 15 seconds) into a compact fixed-voice student trained entirely on synthetic speech. This setting makes pipeline design consequential: teacher errors become training targets, while filtering failed generations can reduce coverage of difficult texts. Thai further introduces challenges from ambiguous word boundaries, lexical tone, names and loanwords, numeric verbalization, and Thai-English code-switching. We study how text preparation, synthetic generation, quality filtering, rejection sampling, and frontend choices affect the resulting student, and where teacher limitations remain. We evaluate CER, Challenge-Set Keyword Accuracy, Prosody Pause Accuracy, speaker similarity, and speaking rate. The resulting 82M-parameter model, Wayu-Paxa-TTS-Edge, enables on-device Thai TTS without reference audio. It achieves 68.2% Challenge-Set Keyword Accuracy (85.5% of Gemini 3.1) and 91.4% pause precision, outperforming its OmniVoice teacher (89.9%) and reaching 94.8% of Gemini 3.1. It also achieves the lowest pause-placement error and intra-word pause rates among the three systems, and 3.7% and 1.1% CER on Thai and English, respectively. We open-source the model and evaluation framework for Thai TTS development.

While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries. We introduce statistically principled early stopping methods that monitor uncertainty signals during generation to mitigate this issue. Our first approach is parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. Our second approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. We conduct empirical evaluations on reasoning tasks across several domains and models. Our results indicate that uncertainty-aware early stopping can improve both efficiency and reliability in LLM reasoning, and we observe especially significant gains for math reasoning.

Ambient AI scribes draft clinical notes under the reassurance that a clinician signs every note. We audited three commercial AI scribes on the same 142 consultations: 565 notes from recorded UK primary-care and US ambulatory encounters plus authored scenarios. Twelve discovery passes proposed 13,678 candidate errors; the 5,898 clearing an importance filter went to an adversarial panel of two models from different families, each told to refute what it could, and 618 survived. One note in three (31.3% [27.0, 35.6]) carries a verified failure, concentrated in allergy and medication information, invented patient identity, and history written up as examination on telephone consultations that can contain none. No product was given a patient record; setting aside the two classes a record would have prefilled, invented identity and dates, the rate is 24.8% [20.8, 29.0]. One failure mode did not fit our scheme, drawn from published scribe-error taxonomies: a treatment the clinician retracts, recorded as delivered care. Two clinicians adjudicated blind, disjoint samples: a physician author upheld 20 of 21 findings (95.2% [77.3, 99.2]) and an independent clinician, not an author, 12 of 12 ([75.8, 100]); both judged every sampled refusal genuine. A failure rate depends on the instrument as much as the scribes. With model, evidence and settings fixed, the review instruction alone moves the share of candidates verified from 9.3% to 79.0%, and the reviewing family moves it too: alone at that instruction the gentler flags 54.8% of notes against 27.8%. Between 28% and 97% of sampled notes carry a failure depending on the standard. Published audits disagree among themselves by a margin instrument differences alone can produce: omission is 54-86% of their errors against our 23.1%. We release all 618 findings with transcript-side evidence, every prompt and model version, and the re-runnable pipeline.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

Large Reasoning Models (LRMs) improve performance by generating explicit intermediate reasoning traces through increased test-time compute, yet the assumption that longer reasoning is consistently beneficial remains under-examined. While recent evidence shows that additional reasoning can lead models to overthink, we ask: "Once a model has reached the correct answer, does further reasoning refine the solution, or deviate from it?" To study the dynamics after correctness, we introduce a prefix-level trajectory evaluation protocol grounded in reasoning sufficiency, defining the minimum reasoning budget required for a model to first generate the correct answer. This allows us to disentangle verbose overthinking, where additional reasoning is redundant but harmless, from harmful overthinking, where continued reasoning destabilizes an already-correct trajectory. Starting from multimodal benchmarks, we find that many instances considered reasoning-intensive require surprisingly little reasoning. Moreover, stopping at the first correct prefix improves accuracy over standard reasoning up to 21%, revealing that current models are limited not only by their ability to reason, but also by their inability to stop at the right time. Furthermore, while common efficiency strategies like early stopping substantially reduce verbose overthinking (up to 50%), they fail to mitigate harmful overthinking. Failure analysis reveals that correctness deviations are mainly driven by logical drift and visual reinterpretation. Finally, we show that our findings generalize to language-only reasoning benchmarks, highlighting harmful overthinking as a broader reliability risk. Code available at https://simonecaldarella.github.io/thinking-past-the-answer.

Forced aligners locate each word of a known transcript in the audio, and many speech pipelines depend on them. In our own data, a widely used CTC aligner inserts a pause at nearly every word boundary, and we found almost no aligner evaluated on telephone audio. We compare nine aligner configurations against human word boundaries on read, conversational and telephone English. The Montreal Forced Aligner (MFA) is the most precise, while CTC aligners consistently start words late, because most frames inside a word go to the blank token. We then fine-tune Whisper with LoRA to predict each word's start and end as timestamp tokens, training only on boundaries placed by people: 47 hours in 87 languages. The resulting aligner is more precise than MFA on every English test condition, and than MFA's own Korean and Dutch models. With FA-Bench's protocol and code on TIMIT, its word-boundary error is 11.6 ms, against 18.2 ms for the best published system. Two choices matter most on telephone conversations: placing the timestamp slots after the word they time, and training on noisy transcripts, which also makes the aligner nearly insensitive to errors in the transcript. The released model reaches 16.0 ms on real landline calls, against 36.5 ms for MFA. The models are available at https://huggingface.co/collections/Scicom-intl/force-alignment

KV-cache eviction caps the memory cost of long reasoning traces but is inherently lossy because the model decodes from a partial view of its history. Under aggressive budgets, this not only lowers accuracy but can also cause runaway degeneration, where the model produces incoherent or repetitive tokens until reaching the length limit. We characterize much of this loss as an information gapf caused by missing context, rather than a capability gap caused by limited model capacity. An evicted 7B model and a full-context 1.5B model make complementary errors, and an oracle choice between their answers recovers 79% of the accuracy gap to the full-KV 7B model. Based on this observation, we propose KV-Rescue, a training-free inference framework that bridges the information gap introduced by KV eviction using a lightweight full-context helper. KV-Rescue interleaves reasoning steps from the two models into a shared trajectory. An online detector uses entropy and compressibility to terminate the generation of incoherent or repetitive base-model candidates early. Across five math benchmarks with Qwen2.5-Math 7B and 72B, KV-Rescue recovers an average of 87% of the accuracy lost to eviction at eviction budget B=64. A decode-cost analysis further shows that preventing runaway degeneration cuts base-model token generation by 43% on average.

Omni-modal large language models (OLMs) redefine human-machine interaction by natively integrating audio, vision, and text. However, existing OLM benchmarks remain anchored to static, accuracy-centric tasks, leaving a critical gap in assessing social interactivity, the fundamental capacity to navigate dynamic cues in natural dialogues. To this end, we propose SocialOmni, a comprehensive benchmark that operationalizes the evaluation of this conversational interactivity across three core dimensions: (i) speaker separation and identification (who is speaking), (ii) interruption timing control (when to interject), and (iii) natural interruption generation (how to phrase the interruption). SocialOmni features 2,000 perception samples and a quality-controlled diagnostic set of 209 interaction-generation instances with strict temporal and contextual constraints, complemented by controlled audio-visual inconsistency scenarios to test model robustness. We benchmarked 12 leading OLMs, which uncovers significant variance in their social-interaction capabilities across models. Furthermore, our analysis reveals a pronounced decoupling between a model's perceptual accuracy and its ability to generate contextually appropriate interruptions, indicating that understanding-centric metrics alone are insufficient to characterize conversational social competence. More encouragingly, these diagnostics from SocialOmni yield actionable signals for bridging the perception-interaction divide in future OLMs.

This position paper argues that the machine learning community should prioritize early-stage quality assurance in annotation pipelines over the prevailing practice of late-stage validation. Data quality bottlenecks increasingly limit foundation model improvement, yet quality assurance research focuses almost exclusively on validation methods rather than validation timing. When validation occurs, not merely what methods are employed, fundamentally determines both error rates and annotation costs. This temporal neglect is puzzling given the well-established "shift-left" principle from software engineering, where empirical studies demonstrate 4--100x cost multipliers for defects detected in later stages (Boehm, 1981; Shull et al., 2002). Annotation pipelines exhibit analogous dynamics: errors caught before annotation begins cost a fraction of those discovered after review cycles complete. We propose a taxonomy of three QA trigger points, namely pre-annotation (T0), post-annotation (T1), and post-review (T2), that decompose annotation workflows into discrete validation opportunities. A parametric error-propagation model formalizes when timing affects final error rates versus only economics, making timing a measurable design variable rather than a configuration afterthought. A survey of 47 recent papers reveals that only 4% report when validation occurs, a striking gap given timing's demonstrated impact in adjacent fields. Without explicit attention to QA timing, the community risks optimizing validation methods while ignoring the structural variable that may matter most. Acting on this position requires three steps: researchers should report QA timing configurations alongside validation methods; annotation platforms should expose timing as a first-class parameter; and the community should run controlled experiments that measure stage-specific detection rates directly.

Large language models (LLMs) now write code in settings where misreading a single word can break safety or cost money, yet we still expect them to overlook stray typos. To probe where useful robustness ends and harmful insensitivity begins, we compile 50 LeetCode problems and craft three minimal prompt perturbations that should vary in importance: (i) progressive underspecification deleting 10 % of words per step; (ii) lexical flip swapping a pivotal quantifier ("max" to "min"); and (iii) jargon inflation replacing a common noun with an obscure technical synonym. Six frontier models, including three "reasoning-tuned" versions, solve each mutated prompt, and their Python outputs are checked against the original test suites to reveal whether they reused the baseline solution or adapted. Among 11 853 generations we observe a sharp double asymmetry. Models remain correct in 85 % of cases even after 90 % of the prompt is missing, showing over-robustness to underspecification, yet only 54 % react to a single quantifier flip that reverses the task, with reasoning-tuned variants even less sensitive than their bases. Jargon edits lie in between, passing through 56 %. Current LLMs thus blur the line between harmless noise and meaning - changing edits, often treating both as ignorable. Masking salient anchors such as function names can force re - evaluation. We advocate evaluation and training protocols that reward differential sensitivity: stay steady under benign noise but adapt - or refuse - when semantics truly change.

We introduce TinyCast, an attention-free zero-shot forecaster that emits a predictive distribution from 146,505 parameters, on the premise that at this size the periodic structure of a context is worth computing rather than learning. A zero-parameter spectral detector supplies the dominant periods, the context is folded on their phase, and a dilated convolutional encoder and a block-autoregressive quantile decoder model the rest. It is smaller than every zero-shot entry on the GIFT-Eval board whose parameter count can be established. On probabilistic accuracy it defines the size-accuracy frontier. Among zero-shot entries declaring no test-data leakage it is the only one below 1.4M parameters that emits a predictive distribution, and every entry scoring better carries at least that budget. On Chronos-ZS and fev-bench every neural model ahead of it carries at least 28 times its parameters. Because the mixing path is convolutions and matrix multiplications only, it exports to static INT8 and forecasts end to end on an embedded device without per-signal fitting.

Plan accuracy alone cannot show whether a speech-delivery decision follows its source: a fixed prior may match the original label yet fail to respond appropriately when a cue changes. VoxReason provides a 100-case verifier benchmark that holds each utterance fixed, edits one designated source-label cue, and scores cited evidence, eight plan fields, and the permitted response. On a source-key-disjoint test of 24 cases, a source-emotion prior reaches plan-slot accuracy 0.958, but none of the 24 edited neutral targets appears in its training labels; its required-change accuracy is 0.000. This diagnoses the support boundary of this prior, not its performance on supported edits. In a complementary 32-case emotion-disjoint test, the prior has seen all edited neutral targets but neither original test emotion; its plan-slot accuracy is 0.219 and required-change accuracy is 1.000. The partitions reuse and overlap the same 100 cases, so these deterministic diagnostics are not independent cohorts or learned-planner results. The benchmark evaluates derived labels and structured plans, not audio input, generated speech, or listener judgments.

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

In physics and engineering literature, the distribution of the excursion-above-zero time distribution (exceedance distribution) for a stationary Gaussian process has been approximated by a stationary switching process with independently distributed switching times. The approach matched the covariance of the clipped Gaussian process with the one for the stationary switching process and the distribution of the latter was used as the so-called independent interval approximation (IIA). The approach successfully assessed the persistency exponent for many physically important processes but left an unanswered question when such an approach leads to a mathematically meaningful and proper exceedance distribution. Here we address this question by proposing an alternative matching of the expected values of the clipped Slepian process and the corresponding switched process initiated at the origin. The method has allowed resolving the mathematical correctness of the matching method for a large subclass of the Gaussian processes with monotonic covariance, for which we provide a sufficient condition for the validity of the IIA. Within this class, the IIA produces a valid distribution for the excursion time and is represented in an explicit stochastic form that connects directly to the covariance of the underlying Gaussian process. We compare the excursion level distributions as well as the corresponding persistency exponents obtained through the IIA method with numerically computed exact distributions, and the simulated distribution for several important Gaussian models. We also argue that for stationary Gaussian processes with a non-monotonic covariance, the IIA fails and should not be used.

Side-channel evaluators routinely inspect leakage tests while acquisition is still running, and extend or stop the campaign based on what they see. Fixed-horizon screening such as the Welch t-test with threshold |t|>4.5 gives no error guarantee for this monitored decision rule. We study anytime-valid leakage detection based on testing by betting: SKIT-type swap-pair e-processes whose false-alarm probability is controlled uniformly over time under an explicit conditional symmetry null. In matched comparisons that share the frozen witness, rows and payoff, first-crossing detection needed 1.68-2.00times the traces of a fixed-horizon randomization test at 80% detection on synthetic streams, and 1.68-2.38times on degraded recordings from an open ML-KEM electromagnetic dataset with the primary Ridge witness at α=0.05. With the same primary witness and level, on undegraded reference and pqm4 recordings the monitored procedure stopped early: its median stopping point was 62-72 and 146-316 evaluation traces, i.e. 2-8% of a conservative 4096-trace budget. Under exact designed nulls on the recorded backgrounds, repeated-look |t|>4.5 screening over all 13000-20000 samples raised a false alarm in 2.7-12.9% of replicates, against 0.0-4.7% for terminal-only screening and no rejection by a sample-wise e-Bonferroni process, which in a prespecified follow-up detected natural-label associations in 4 of 4 backgrounds after 840-3288 traces. All recordings come from one device, and natural-label results are descriptive; we state the assumptions each claim requires.

In mobile telecommunications, alarms act as early warning signals. They are triggered when a cell, the basic unit of radio coverage, shuts down or behaves abnormally. This signals a degradation in service quality, which directly affects the customer experience. To fix the issue, operators rely on preset thresholds to decide when an engineer should be sent out. In practice, these thresholds are set manually and remain fixed regardless of the time of day, traffic levels, or overall network conditions. This often leads to serious faults slipping through during busy hours, while minor issues can cause unnecessary callouts when the network is quiet. This paper presents a machine learning framework that automatically predicts four alarm thresholds, audit window duration, inactive time limit, total fluctuation count, and per hour fluctuation limit, from live network behavior. Since no ground truth labels exist for thresholds, we introduce a percentile guided label derivation strategy and evaluate four models on an anonymized dataset of 10,648 cells across three vendors and nine regions from a real 4G network, comprising a Gradient Boosted Trees baseline, a CNN-BiLSTM with attention, the proposed PCTN, and an iTransformer. PCTN performs the best overall with respect to three of the four targets, outperforming a state-of-the-art iTransformer while using 83 percent fewer parameters. Its mixed output heads and dynamic alpha mechanism produce thresholds that are both accurate and interpretable, allowing operators to inspect and adjust the learned policy without retraining. All comparisons are statistically significant at p < 0.001. The framework undergoes daily retraining using new data, which enables the thresholds to constantly adjust to changes in the network.

As AI agents are increasingly deployed in real personal and corporate settings (email accounts, development workflows, company databases, etc.), safety considerations surrounding these agents become paramount. Although much work has focused on agent safety in the presence of an adversary, we show that agents can exhibit misaligned behavior even in benign settings, taking unsafe actions when those actions are instrumental to task completion. We study this failure mode through the lens of corrigibility, the safety desideratum that agents remain amenable to human correction, interruption, or shutdown. To demonstrate this tendency, we introduce a benchmark in which agents are asked to complete realistic, computer-use tasks but are confronted with a corrigibility obstacle: a human interrupt, a login page, or a shutdown notification. We then evaluate whether agents choose to violate corrigibility in order to complete the task -- overriding the human, accessing private passwords, rewiring shutdown. We find that the overwhelming majority of frontier models tested frequently bypass user interruptions or restrictions. In addition, better model performance appears to lead to greater misalignment. Finally, even when models are completely corrigible initially, we show there are no guarantees that the subagents they create are. Our work highlights the critical need for principled, corrigibility-focused alignment methods in autonomous agents.

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics. A 661.6M parameter model (approx. 65% code/technical text; no dedicated SFT) and a 1,109M model (web-heavy multi-phase curriculum; 6B-token tool-SFT) share decoder, tokenizer, and special tokens, scoring almost identically on lenient tool-use metrics (B4: 0.660 vs. 0.650). Verbatim-reproduction checks on training examples separate them completely: the 600M emits valid tool calls with generalized arguments on 6/6 examples; the 1B does so on 0/6 across checkpoints. A first-token probe localizes the 1B's failure to a missing prior (prob. 10^{-4}--10^{-5} on <|tool_call|>), which was erased by its web-heavy training phase. A targeted SFT recipe (diverse corpus, 5x higher learning rate, 2,202 steps, ~3.3 GPU-hours) repairs the 1B using three orders of magnitude fewer tokens than the failed phase. On all 269 corpus rows, valid emission rises from 0.100 to 0.959 (600M: 0.926). On 238 unseen prompts, the repaired 1B passes 0.536 vs. the 600M's 0.428 (p = 0.004). Embedding-drift checks show the repair did not move the trigger token's tied embedding (97.7% of the bf16 table remains bit-identical), meaning changes live in the surrounding network. Both models over-trigger, rarely answering negative prompts without a call (0.09 for 600M, 0.17 for repaired 1B). Factorial analyses confirm all repair configurations install the format, though suppression benefits from a diverse corpus remain a hypothesis due to seed sensitivity. This cheap diagnostic ladder costs minutes of CPU time and should gate tool-use claims on small models.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration such as evaluation batch size, GPU count, and GPU version can introduce significant difference in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision -- while critical for reproducibility -- is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

In interactions between users and language model agents, user utterances frequently exhibit ellipsis (omission of words or phrases) or imprecision (lack of exactness) to prioritize efficiency. This can lead to varying interpretations of the same input based on different assumptions or background knowledge. It is thus crucial for agents to adeptly handle the inherent ambiguity in queries to ensure reliability. However, even state-of-the-art large language models (LLMs) still face challenges in such scenarios, primarily due to the following hurdles: (1) LLMs are not explicitly trained to deal with ambiguous utterances; (2) the degree of ambiguity perceived by the LLMs may vary depending on the possessed knowledge. To address these issues, we propose Alignment with Perceived Ambiguity (APA), a novel pipeline that aligns LLMs to manage ambiguous queries by leveraging their own assessment of ambiguity (i.e., perceived ambiguity). Experimental results on question-answering datasets demonstrate that APA empowers LLMs to explicitly detect and manage ambiguous queries while retaining the ability to answer clear questions. Furthermore, our finding proves that APA excels beyond training with gold-standard labels, especially in out-of-distribution scenarios.

Closed-source frontier labs do not disclose parameter counts. Storing F facts requires at least F/(bits per parameter) weights, so factual recall lower-bounds parameter count--an intrinsic, serving-independent signal, though (as we show) a coarse one. We introduce Incompressible Knowledge Probes (IKPs), a benchmark of 1,400 factual questions spanning 7 tiers of obscurity, designed to isolate knowledge that cannot be derived by reasoning or compressed by architectural improvements. We score with no hallucination penalty (lambda = 0: IKP accuracy is simply the fraction of probed facts answered correctly), which removes both the penalty hyperparameter and the per-tier flooring choice; a full lambda x flooring ablation shows the calibration is robust across scoring choices while individual estimates are not, motivating the no-penalty default. We calibrate a log-linear mapping from IKP accuracy to parameter count on 93 open-weight models (135M-1,600B) spanning 19 vendors, achieving R^2 = 0.910; leave-one-out cross-validation confirms generalization (median fold error 1.48x, 72% within 2x, 86% within 3x). The instrument is deliberately coarse--its 90% prediction interval spans ~3x in either direction, wider than inference economics--so IKP recovers order-of-magnitude effective capacity and relative rankings, not precise parameter counts. For Mixture-of-Experts models, total parameters predict knowledge (R^2 = 0.67) better than active parameters (R^2 = 0.41). We evaluate 201 models from 27 vendors on a curated probe set (1,311 of 1,400 probes surviving name-collision and label-ambiguity filters) and report effective knowledge capacity for all major proprietary frontier models as prediction bands rather than point estimates; for heavily safety-tuned models these are lower bounds, since refusal policy can suppress tens of percentage points of otherwise-answerable capacity.

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

The quality with which language models follow instructions falls as the load grows, yet there is still no quantity that expresses the limit of a model in units of load with an uncertainty. We propose an instrument in which the load consists of rules that replace single-token words in a text without meaning: such an instruction is atomic, the rules are independent by the construction of the task, and the outcome of every application of a rule is determined by a computable reference answer. Every occurrence of a word to be replaced becomes a Bernoulli trial with a known position in the answer and a known rule number. The statistical model first measures the shape of the dependence of the probability of a right replacement on the text length, the word position, the replacement density, the number of rules and the rule number, then describes it by smooth laws and restores a reference level that does not depend on the choice of the coding. The model yields the instruction capacity - the number of rules held at a given share of right replacements - with an uncertainty. A measurement of GPT-4o mini (2958 requests, 642,430 trials) gave regular laws for every load factor and for the collapse of the answer; the instruction capacity was about 37-38 rules at a share of right replacements of 0.95, and about 18-19 rules at the full replacement density. The code of the instrument, the measurement data and the passport are open, and every number of the passport is recomputed from the published points. The passport of a model serves as the calibration of a future meter of the instruction load of prompts.

Synthetic data generation has become a cornerstone for advancing large language models. However, the lack of the quantitative analysis for error tolerance became a critical bottleneck. Consequently, current filtering strategies fluctuate between two extremes: they are either overly aggressive, risking the exclusion of potentially valuable samples, or overly permissive, failing to eliminate erroneous samples effectively. To bridge this gap, this paper introduces Atomic Tree Operation Modeling (ATOM), a framework that decomposes data into functional units (f(x)rightarrow y). ATOM distinguishes benign Operand x perturbations from fatal Operator f perturbations. The former are needlessly discarded by aggressive filtering, while the latter slip through permissive filtering. Our experiments reveal a double dissociation: models are robust to operand perturbations but collapse under operator perturbations. By prioritizing operator over aggressive operand precision, our ATOM-synthesized data outperforms rigorous baselines (e.g., +3.1% gain over LIMA), suggesting that operator diversity matters more than operand precision. Our code is available at https://github.com/Lut-hub/ATOM.

Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.

We describe our system for LitTraceQA (GroundLM @ EMNLP 2026): given a research question, retrieve the relevant papers from a pool of 27,487, cite the page and the table or figure where the answer lives, and answer in a requested format. Our main finding is that evidence grounding is limited by retrieval, not by reading. The page selector put the annotator's page, which we call the gold page, in front of the model that locates evidence only about half the time (52.6% gold-page recall), while that model, given the page, cited the right one in 45 of the 48 locators it emitted (94%). When the page was missing it rarely said so: of 45 such cases it returned nothing 14 times, a wrong page 24 times, and a correct page 7 times, so the pipeline failed quietly almost twice as often as it failed visibly. Since the failure was that the right page was never shown, the fix is to stop choosing: each retrieved paper fits in the model's context, so we show it whole. Page ranking survives only as a fallback inside papers too long to fit, which no test-split paper was, and gold-page recall reaches 100% on the papers we can parse. Separately, questions that identify their target by position rather than content, such as "the first author of the 24th reference", are served by parsing rather than retrieval: we resolve the bibliography into an addressable list, which also supplies identifiers the evidence metric scores. The final system scores 0.762 paper F_1, 0.441 evidence F_1 and 0.920 multiple-choice accuracy on the held-out test split. Because the pipeline depends on a closed model without seed control, we release a harness that verifies the paper's central claims against committed artifacts.

Stochastic rounding (SR) is a probabilistic method used to round numbers to floating-point and fixed-point representations. In length n summation, the worst-case error of SR grows as n with high probability, unlike for standard modes, like round-to-nearest (RN), which grows as n. For this reason, the former is increasingly employed in large-scale, low-precision computations as an RN alternative. Additionally, SR alleviates stagnation, whereby relatively small summands are completely rounded off and do not contribute to the sum. We provide an update to [Croci et al., Roy. Soc. Open Sci. 9.3 (2022), pp. 1-25], a survey which discusses the development and use of SR between 1949 and 2022, citing over 100 references. Since then, there has been a surge of new research, and this update covers almost four years of further progress in applying, analysing, and implementing SR. Our main focus is limited-precision stochastic rounding, a new variant that fixes the precision of the random numbers used. We provide insights into industrial and numerical analysis activities surrounding SR, highlighting the next possible steps in making this rounding mode more widely available in hardware.

Production-grade LLM systems require robust adherence to dozens or even hundreds of instructions simultaneously. However, the instruction-following capabilities of LLMs at high instruction densities have not yet been characterized, as existing benchmarks only evaluate models on tasks with a single or few instructions. We introduce IFScale, a simple benchmark of 500 keyword-inclusion instructions for a business report writing task to measure how instruction-following performance degrades as instruction density increases. We evaluate 20 state-of-the-art models across seven major providers and find that even the best frontier models only achieve 68% accuracy at the max density of 500 instructions. Our analysis reveals model size and reasoning capability to correlate with 3 distinct performance degradation patterns, bias towards earlier instructions, and distinct categories of instruction-following errors. Our insights can help inform design of instruction-dense prompts in real-world applications and highlight important performance-latency tradeoffs. We open-source the benchmark and all results for further analysis at https://distylai.github.io/IFScale.

Current TTS systems typically rely on open-loop, single-pass generation and can produce sporadic local prosodic defects, such as misplaced stress, unnatural pauses, or flattened intonation, that utterance-level metrics often fail to expose. We present LoopTTS, a judge-guided Filter-Judge-Refiner framework for recovering low-quality TTS outputs diagnosed by an AudioLLM. Given an initial utterance from a base TTS model, an AudioLLM Judge identifies salient prosodic issues and generates structured refine instructions; a Refiner, our fine-grained instruction-following TTS model, then performs guided expressive re-synthesis conditioned on the initial utterance, target text, and instruction. To train the Refiner, we construct Refiner-DB, a 42K-example AudioLLM-annotated dataset with word-level prosodic weak supervision. Human evaluation on diagnosed low-quality utterances shows that LoopTTS can detect perceptually salient errors and correct them with the Refiner, outperforming raw generated audio and practical open-loop re-generation baselines in recovery quality. The Refiner also demonstrates stronger instruction-following ability for stress and pause control in targeted prosody modification.

Monitoring complex industrial assets relies on engineer-authored symbolic rules that trigger based on sensor conditions and prompt technicians to perform corrective actions. The bottleneck is not detection but response: translating rules into maintenance steps requires asset-specific knowledge gained through years of practice. We investigate whether LLMs can serve as decision support for this rule-to-action step and introduce , a benchmark of 6{,}690 expert-validated multiple-choice questions from 118 rule-action pairs across 16 asset types. We contribute (i) a symbolic-to-MCQA pipeline normalizing rules to Disjunctive Normal Form with embedding-based distractor sampling, (ii) five variants probing distinct failure modes (Pro, Pert, Verbose, Aug, Rationale), and (iii) a benchmark of 29 LLMs and 4 embedding baselines. A human evaluation (9 practitioners, mean 45.0\%) confirms requires specialist knowledge beyond operational experience. Three findings stand out. The frontier has closed: the top three LLMs lie within one Macro point, with Bradley-Terry Elo placing claude-opus-4-6 30 points above the next model. Yet \,Pro exposes brittleness, with every model losing 13--60\% relative accuracy under distractor expansion. \,Aug exposes pattern-matching: under condition inversion, frontier models still select the original answer 49--63\% of the time. The deployment bottleneck is not capability but calibration: frontier models handle template-style fault detection but break under structural perturbation.

As large language models are increasingly deployed in retrieval-augmented generation and agentic systems that accumulate extensive context, understanding how distracting information affects long-context performance becomes critical. Prior work shows that semantically relevant yet misleading documents degrade performance, but the quantitative relationship between the proportion of distractors and performance remains unstudied. In this work, we systematically vary the hard-distractor proportion in fixed-length contexts, revealing a striking nonlinear pattern: as the proportion of hard distractors increases, performance drops sharply within the first small fraction, while the remainder of the range yields only marginal additional decline. We term this ''The First Drop of Ink'' effect, analogous to how a single drop of ink contaminates water. Our theoretical and empirical analyses grounded in attention mechanics show that hard distractors capture disproportionate attention even at small proportions, with diminishing marginal impact as their proportion grows. Controlled experiments further show that filtering gains mainly come from context-length reduction rather than distractor removal; substantial recovery requires reducing the hard-distractor proportion to near zero, highlighting the importance of upstream retrieval precision.

Spoken dialogue systems powered by large language models have demonstrated remarkable abilities in understanding human speech and generating appropriate spoken responses. However, these systems struggle with end-turn detection (ETD) -- the ability to distinguish between user turn completion and hesitation. This limitation often leads to premature or delayed responses, disrupting the flow of spoken conversations. In this paper, we introduce the ETD Dataset, the first public dataset for end-turn detection. The ETD dataset consists of both synthetic speech data generated with text-to-speech models and real-world speech data collected from web sources. We also propose SpeculativeETD, a novel collaborative inference framework that balances efficiency and accuracy to improve real-time ETD in resource-constrained environments. Our approach jointly employs a lightweight GRU-based model, which rapidly detects the non-speaking units in real-time on local devices, and a high-performance Wav2vec-based model running on the server to make a more challenging classification of distinguishing turn ends from mere pauses. Experiments demonstrate that the proposed SpeculativeETD significantly improves ETD accuracy while keeping the required computations low. Datasets and code will be available after the review.

As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that confidence ranking largely breaks at atomic resolution, even when clearer separability emerges at coarser line-level units. SALT further enables controlled atom-level interventions throughout generation, revealing two separable drivers of future errors: propagation from corrupted prefixes, dominated by global context correctness, and bounded degradation from increasing answer-context length. Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.

Low-precision Transformer systems increasingly quantize attention matrix multiplications, while softmax often remains at higher precision. During pretraining, an approximate softmax changes the gradients that train the model as well as its forward computation. We study this interaction with K-interval attention, which approximates the exponential using K+1 grid values. We vary per-row grid calibration, interpolation versus hard rounding, and the placement of a straight-through surrogate relative to normalization. We derive the corresponding backward rules, including calibration derivatives, and compare these choices in pretraining experiments matched on model, data, and optimizer. Detaching the row extrema leaves the forward computation unchanged but produces a delayed increase in validation loss. With hard rounding at K=4, min-max calibration and a pre-normalization surrogate incur a large loss gap; changing either choice substantially reduces it. At 124M parameters and 2.5B training tokens, fixed-window calibration with a post-normalization surrogate yields a validation loss gap of +0.019 nats relative to softmax at K=4, and with a pre-normalization surrogate yields +0.004 nats at K=16.

In speech enhancement and source separation, signal-to-noise ratio is a ubiquitous objective measure of denoising/separation quality. A decade ago, the BSS_eval toolkit was developed to give researchers worldwide a way to evaluate the quality of their algorithms in a simple, fair, and hopefully insightful way: it attempted to account for channel variations, and to not only evaluate the total distortion in the estimated signal but also split it in terms of various factors such as remaining interference, newly added artifacts, and channel errors. In recent years, hundreds of papers have been relying on this toolkit to evaluate their proposed methods and compare them to previous works, often arguing that differences on the order of 0.1 dB proved the effectiveness of a method over others. We argue here that the signal-to-distortion ratio (SDR) implemented in the BSS_eval toolkit has generally been improperly used and abused, especially in the case of single-channel separation, resulting in misleading results. We propose to use a slightly modified definition, resulting in a simpler, more robust measure, called scale-invariant SDR (SI-SDR). We present various examples of critical failure of the original SDR that SI-SDR overcomes.

In real-world Tool-Integrated Reasoning (TIR) scenarios, where LLMs interleave reasoning with external tool calls, a major source of inefficiency is that the toolcalls create pauses between LLM requests and cause KV-Cache eviction, forcing recomputation. Also, the long, unfiltered response returned by external tools inflates the KV-Cache, so each decode step spends more time loading the growing cache and thus becomes steadily slower as context length increases. However, existing efficiency metrics like token counts and toolcall counts fail to capture the real model inference latency. To address this, we introduce PTE (Prefill Token Equivalents), a hardware-aware TIR-efficiency metric that unifies internal reasoning and external tool-use costs while explicitly accounting for non-reusable KV-Cache and long-tool-response scenarios. Validation in a high-concurrency industrial setting indicates that PTE aligns significantly better with wall-clock latency than standard token counts, while maintaining consistent efficiency rankings across diverse hardware profiles. We conduct extensive experiments across five TIR benchmarks, quantify their PTE costs, and identify four inefficiency patterns that appear in TIR. We also discover that trajectories with higher PTE costs tend to have lower reasoning correctness, indicating that simply using more tools does not improve the quality of the answer.

Reasoning failures in large language models (LLMs) are typically measured only at the end of a generation, yet many failures manifest as a process-level breakdown: the model "loses the thread" mid-reasoning. We study whether such breakdowns are detectable from inference-time observables available in standard APIs (token log probabilities), without any training or fine-tuning. We define a simple instability signal that combines consecutive-step distributional shift (JSD) and uncertainty (entropy), summarize each trace by its peak instability strength, and show that this signal reliably predicts failure. Across GSM8K and HotpotQA, instability strength predicts wrong answers with above-chance AUC and yields monotonic bucket-level accuracy decline at scale across model sizes. Crucially, we show that instability is not uniformly harmful: early instability can reflect subsequent stabilization and a correct final answer (corrective instability), whereas late instability is more often followed by failure (destructive instability), even at comparable peak magnitudes, indicating that recoverability depends not only on how strongly the distribution changes but also on when such changes occur relative to the remaining decoding horizon. The method is model-agnostic, training-free, and reproducible, and is presented as a diagnostic lens rather than a corrective or control mechanism.

Benchmarks can run without determining what their results license. We freeze a large evaluation collection and attempt to replay its historical claims. Most units stop because the evidence required for replay is not bound. Where replay is possible, different claims remain stable at different resolutions. We make this otherwise implicit inference step explicit and executable.

Medication reconciliation at clinical handoffs is a high-stakes, error-prone process. Large language models are increasingly proposed to assist with this task using FHIR-structured patient records, but a fundamental and largely unstudied variable is how the FHIR data is serialised before being passed to the model. We present the first systematic comparison of four FHIR serialisation strategies (Raw JSON, Markdown Table, Clinical Narrative, and Chronological Timeline) across five open-weight models (Phi-3.5-mini, Mistral-7B, BioMistral-7B, Llama-3.1-8B, Llama-3.3-70B) on a controlled benchmark of 200 synthetic patients, totalling 4,000 inference runs. We find that serialisation strategy has a large, statistically significant effect on performance for models up to 8B parameters: Clinical Narrative outperforms Raw JSON by up to 19 F1 points for Mistral-7B (r = 0.617, p < 10^{-10}). This advantage reverses at 70B, where Raw JSON achieves the best mean F1 of 0.9956. In all 20 model and strategy combinations, mean precision exceeds mean recall: omission is the dominant failure mode, with models more often missing an active medication than fabricating one, which changes how clinical safety auditing priorities should be set. Smaller models plateau at roughly 7-10 concurrent active medications, leaving polypharmacy patients, the patients most at risk from reconciliation errors, systematically underserved. BioMistral-7B, a domain-pretrained model without instruction tuning, produces zero usable output in all conditions, showing that domain pretraining alone is not sufficient for structured extraction. These results offer practical, evidence-based format recommendations for clinical LLM deployment: Clinical Narrative for models up to 8B, Raw JSON for 70B and above. The complete pipeline is reproducible on open-source tools running on an AWS g6e.xlarge instance (NVIDIA L40S, 48 GB VRAM).

Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards. Recent work agrees these checkers are weak and patches them by hand---extra input distributions, fuzzing recipes, tighter tolerances---with no way to measure whether any patch suffices. We introduce mutation analysis as an adequacy metric for kernel-benchmark oracles: deterministic rules inject 10{,}303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7{,}384 of them with an independent kill witness; any test protocol is scored by the fraction it detects. The official check misses one in six witnessed faults (16.9%), deterministically, and the misses are skewed by family: 8.7% of arithmetic faults escape, but 78.6% of precision faults do. The metric explains why (a tolerance blind band growing with reduction size; a measured ceiling on input aggressiveness set by legitimate floating-point variance), audits the strongest existing patch (KernelBench-Verified's gain splits into +4.0 points from hidden inputs and +4.5 from tighter tolerance, a split its authors could not compute), and exposes a published fuzzing recipe that rejects correct kernels 107 times. Optimizing suites over the kill matrix reaches 98.0% detection with two inputs per problem (94.8% held-out), and the measurement's fault taxonomy teaches a test generator more than the raw faults themselves. Across 48 whole architectures, the blindness grows with scale, concentrating in deep homogeneous pipelines, and two problems prove unrefereeable: their official references violate the benchmark's own tolerance against fp64. We release everything as https://huggingface.co/datasets/Elfsong/KernelBench-M{KernelBench-M}.

As agent systems become more widely used, multiple agent sessions increasingly run alongside pre-existing user tasks in the same environment, sharing resources with limited capacity or mutually exclusive states. This creates a safety risk: when granted sufficient privileges, an agent may resolve a resource conflict by terminating or otherwise disrupting an existing task rather than reporting it. In this work, we identify and formalize this failure mode, which we term destructive resource preemption: obtaining the resources required for a requested task by terminating, overwriting, evicting, or degrading an incumbent task. To systematically study this risk, we introduce ClashBench, an executable benchmark comprising 268 validated conflict cases across 55 resource types, and evaluate 17 models through Codex, Claude Code, and OpenCode. We observe destructive preemption in 44.5% of trajectories, where the agent completes the requested task while causing the incumbent task to fail its health check. We also show that prompt-based safeguards are insufficient: an instruction to avoid affecting existing tasks reduces but does not eliminate preemption, while an instruction explicitly authorizing the agent to stop local processes increases it. More concerningly, in 31.9% of successful destructive-preemption cases, the final response mentions neither the resource conflict nor the action taken to resolve it, raising concerns about possible concealment. These findings establish destructive resource preemption as a broad safety risk in privileged agent systems and motivate stronger privilege controls, task isolation, and conflict-aware safeguards.

Retrieval-augmented QA pipelines often route retrieved passages through an LLM rewriter before a smaller reader, lifting F1 by tens of points on multi-hop benchmarks; this gain is typically credited to improved evidence quality. We ask whether that lift is causally driven by the gold answer string appearing in the rewritten context rather than by curation per se, using a controlled intervention audit. For each rewritten context we re-run the reader after one of four controlled edits to the compile output: removing the gold answer span, replacing a length-matched random non-answer span (placebo), or injecting the gold into rewrites where it was absent (at the prefix or at a midpoint sentence boundary). Across twelve completed (cell, baseline) intervention runs spanning three reader families (Qwen2.5-7B, Qwen3.5-35B, GLM-4.7), two datasets (HotpotQA, 2WikiMultihopQA), and three compiler arrangements (MA-only, MB-only, MA+verify), removing the gold answer drops reader F1 by 28 to 64 points beyond the length-matched placebo on paired answer-in-compile strata, and prepending the gold into rewrites that lacked it raises F1 by +0.7 to +9.7 points in 10 of 12 (cell, baseline) combinations. A companion five-sentinel audit shows the conventional single-[MASK] probe is itself sentinel-fragile: on 2Wiki it reports a +4.12~F1 ``non-leakage residual'' that flips to -3.33 to -7.81~F1 under four alternative sentinels and fails an equivalence test for three of those four (1/4~pass). We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard.

Large Language Models (LLMs) are highly sensitive to subtle, non-semantic variations in prompt phrasing and formatting. In this work, we present the first systematic evaluation of 5 methods for improving prompt robustness within a unified experimental framework. We benchmark these techniques on 8 models from Llama, Qwen and Gemma families across 52 tasks from Natural Instructions dataset. Our evaluation covers robustness methods from both fine-tuned and in-context learning paradigms, and tests their generalization against multiple types of distribution shifts. Finally, we extend our analysis to GPT-4.1 and DeepSeek V3 to assess frontier models' current robustness to format perturbations. Our findings offer actionable insights into the relative effectiveness of these robustness methods, enabling practitioners to make informed decisions when aiming for stable and reliable LLM performance in real-world applications. Code: https://github.com/AIRI-Institute/when-punctuation-matters.

We investigate a failure mode of large language models (LLMs) in which plain-text prompts elicit excessive outputs, a phenomenon we term Overflow. Unlike jailbreaks or prompt injection, Overflow arises under ordinary interaction settings and can lead to elevated serving cost, latency, and cross-user performance degradation, particularly when scaled across many requests. Beyond usability, the stakes are economic and environmental: unnecessary tokens increase per-request cost and energy consumption, compounding into substantial operational spend and carbon footprint at scale. Moreover, Overflow represents a practical vector for compute amplification and service degradation in shared environments. We introduce BenchOverflow, a model-agnostic benchmark of nine plain-text prompting strategies that amplify output volume without adversarial suffixes or policy circumvention. Using a standardized protocol with a fixed budget of 5000 new tokens, we evaluate nine open- and closed-source models and observe pronounced rightward shifts and heavy tails in length distributions. Cap-saturation rates (CSR@1k/3k/5k) and empirical cumulative distribution functions (ECDFs) quantify tail risk; within-prompt variance and cross-model correlations show that Overflow is broadly reproducible yet heterogeneous across families and attack vectors. A lightweight mitigation-a fixed conciseness reminder-attenuates right tails and lowers CSR for all strategies across the majority of models. Our findings position length control as a measurable reliability, cost, and sustainability concern rather than a stylistic quirk. By enabling standardized comparison of length-control robustness across models, BenchOverflow provides a practical basis for selecting deployments that minimize resource waste and operating expense, and for evaluating defenses that curb compute amplification without eroding task performance.

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 12 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 10 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information.

The needle-in-a-haystack (NIAH) test, which examines the ability to retrieve a piece of information (the "needle") from long distractor texts (the "haystack"), has been widely adopted to evaluate long-context language models (LMs). However, this simple retrieval-based test is indicative of only a superficial form of long-context understanding. To provide a more comprehensive evaluation of long-context LMs, we create a new synthetic benchmark RULER with flexible configurations for customized sequence length and task complexity. RULER expands upon the vanilla NIAH test to encompass variations with diverse types and quantities of needles. Moreover, RULER introduces new task categories multi-hop tracing and aggregation to test behaviors beyond searching from context. We evaluate ten long-context LMs with 13 representative tasks in RULER. Despite achieving nearly perfect accuracy in the vanilla NIAH test, all models exhibit large performance drops as the context length increases. While these models all claim context sizes of 32K tokens or greater, only four models (GPT-4, Command-R, Yi-34B, and Mixtral) can maintain satisfactory performance at the length of 32K. Our analysis of Yi-34B, which supports context length of 200K, reveals large room for improvement as we increase input length and task complexity. We open source RULER to spur comprehensive evaluation of long-context LMs.

End-to-end (E2E) spoken dialogue systems are increasingly replacing cascaded pipelines for voice-based human-AI interaction, processing raw audio directly without intermediate transcription. Existing benchmarks primarily evaluate these models on synthetic speech and single-turn tasks, leaving realistic multi-turn conversational ability underexplored. We introduce Audio MultiChallenge, an open-source benchmark to evaluate E2E spoken dialogue systems under natural multi-turn interaction patterns. Building on the text-based MultiChallenge framework, which evaluates Inference Memory, Instruction Retention, and Self Coherence, we introduce a new axis Voice Editing that tests robustness to mid-utterance speech repairs and backtracking. We further augment each axis to the audio modality, such as introducing Audio-Cue challenges for Inference Memory that require recalling ambient sounds and paralinguistic signals beyond semantic content. We curate 452 conversations from 47 speakers with 1,712 instance-specific rubrics through a hybrid audio-native agentic and human-in-the-loop pipeline that exposes model failures at scale while preserving natural disfluencies found in unscripted human speech. Our evaluation of proprietary and open-source models reveals that even frontier models struggle on our benchmark, with Gemini 3 Pro Preview (Thinking), our highest-performing model achieving a 54.65% pass rate. Error analysis shows that models fail most often on our new axes and that Self Coherence degrades with longer audio context. These failures reflect difficulty of tracking edits, audio cues, and long-range context in natural spoken dialogue. Audio MultiChallenge provides a reproducible testbed to quantify them and drive improvements in audio-native multi-turn interaction capability.

Natural turn-taking in full-duplex voice interaction requires determining from partial speech whether a pause reflects hesitation or a completed conversational intent. Acoustic voice activity detection lacks this semantic information, while cascaded ASR-based endpointing introduces transcription dependence and additional processing stages. We formulate semantic endpoint detection as a causal audio-language reasoning task and introduce FD-VAD, an ASR-free streaming endpointer that maps bounded causal audio windows directly to Continue/Stop decisions. FD-VAD combines a frozen speech encoder with a lightweight modality adapter and a parameter-efficiently adapted language model, using a last-chunk training objective for streaming inference. We further introduce confidence-gated endpoint commitment to control interruption versus delay and boundary-focused hard-negative sampling to improve decisions around ambiguous turn boundaries. Across in-domain and conversational evaluations, FD-VAD outperforms strong streaming and non-streaming semantic turn classifiers, and achieves the highest EOT recall among qualifying systems on TurnBench dev set 0.853 (at FP<=0.10) in a zero-shot setting. These results show that semantic endpointing can be performed directly from streaming audio without intermediate ASR or dialogue state tracking.

Tool-augmented LLM agents tend to call tools indiscriminately, even when the model can answer directly. Each unnecessary call wastes API fees and latency, yet no existing benchmark systematically studies when a tool call is actually needed. We propose When2Tool, a benchmark of 18 environments (15 single-hop, 3 multi-hop) spanning three categories of tool necessity -- computational scale, knowledge boundaries, and execution reliability -- each with controlled difficulty levels that create a clear decision boundary between tool-necessary and tool-unnecessary tasks. We evaluate two families of training-free baselines: Prompt-only (varying the prompt to discourage unnecessary calls) and Reason-then-Act (requiring the model to reason about tool necessity before acting). Both provide limited control: Prompt-only suppresses necessary calls alongside unnecessary ones, and Reason-then-Act still incurs a disproportionate accuracy cost on hard tasks. To understand why these baselines fail, we probe the models' hidden states and find that tool necessity is linearly decodable from the pre-generation representation with AUROC 0.89--0.96 across six models, substantially exceeding the model's own verbalized reasoning. This reveals that models already know when tools are needed, but fail to act on this knowledge during generation. Building on this finding, we propose Probe&Prefill, which uses a lightweight linear probe to read the hidden-state signal and prefills the model's response with a steering sentence. Across all models tested, Probe&Prefill reduces tool calls by 48% with only 1.7% accuracy loss, while the best baseline at comparable accuracy only reduces 6% of tool calls, or achieves a similar tool call reduction but incurs a 5times higher accuracy loss. Our code is available at https://github.com/Trustworthy-ML-Lab/when2tool

We present our submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, which ranked first in the large-model division and second in the <=2B division. The task requires a wearable assistant to decide after each eight-second segment of egocentric video whether to intervene or remain silent. Our approach has two main components. First, we reformulate intervention timing as single-token classification. Rather than generating either interrupt<utterance> or silent, the model predicts yes or no, and we derive the decision from the renormalised probabilities of these two tokens. This formulation improved macro-F1 by 0.249 and G-mean by 0.30 over free-form generation. Second, because labelled data were limited to the released validation set, we generated additional supervision using a tool-calling video agent that inspects each clip and assigns intervention timestamps. A narration-only alternative was four times larger and ten times cheaper, but transferred worse than supervision from an unrelated real corpus, suggesting that visual grounding is more important than annotation volume for this task.

Real-time interaction models -- Moshi, MiniCPM-o, Qwen-Omni -- turn serving into a periodic real-time task: on every frame a session ingests streaming audio and must respond by a recurring wall-clock deadline, while its KV cache grows monotonically and stays pinned for the whole conversation. This regime hides a dangerous failure mode. On a real full-duplex stack, sustained load does not degrade serving gracefully: it falls off a cliff, jumping in one step from milliseconds per frame to a stalled engine when accumulated session state exhausts the KV pool. The collapse is metastable -- identical five-minute runs collapse or survive on run-to-run variance -- and silent: latency and deadline-miss metrics read healthy throughout. We show one move restores both stability and observability: bound each session's resident state, and latency starts telling the truth. Metronome's in-engine KV window eliminates the collapse (0/20 vs. 14/20 runs across two batches) and turns per-frame latency into a monotone load signal, on which an online admission controller discovers the schedulable concurrency; without the window, the identical controller over-admits into the wall. A first-order model predicts the collapse time within a few percent on the headline model, and a quality probe validates the bound's design by ablation: the window alone is quality-free in turn-based decoding, and its few pinned attention-sink tokens are what keep free-running generation healthy. Everything is measured end-to-end on real audio, across four interaction models on one GPU.

Recent advancements in large language models (LLMs) have shown their remarkable capacities in many NLP tasks. However, their substantial size often presents challenges for deployment. This necessitates efficient techniques for model compression, with quantization emerging as a prominent solution. Despite its benefits, the exact impact of quantization (from 2- to 6-bit) on the performance and accuracy of LLMs remains an active area of research. This paper investigates the performance of eight LLMs at various quantization levels, focusing on tasks such as MMLU-Pro for knowledge processing and reasoning, CRUXEval for code comprehension, and MuSR for reading comprehension. Our results show a consistent trend where higher precision (e.g., 8-bit Q8\_0) yields improved performance, albeit with diminishing returns. Aggressive quantization (e.g., 2-bit Q2\_K) usually retains acceptable accuracy, though some models show a substantial loss in performance. Our findings indicate that while lower bit precision generally reduces performance, the impact varies across models and tasks. Larger models show greater resilience to aggressive quantization, but can still undergo significant drops at lower precision levels. Mid-sized models in the 7-9 billion parameter range strike an optimal balance between efficiency and resource usage. Such results provide insights into the trade-offs between model size, quantization, and performance.

The pursuit of computational efficiency has driven the adoption of low-precision formats for training transformer models. However, this progress is often hindered by notorious training instabilities. This paper provides the first mechanistic explanation for a long-standing and unresolved failure case where training with flash attention in low-precision settings leads to catastrophic loss explosions. Our in-depth analysis reveals that the failure is not a random artifact but caused by two intertwined phenomena: the emergence of similar low-rank representations within the attention mechanism and the compounding effect of biased rounding errors inherent in low-precision arithmetic. We demonstrate how these factors create a vicious cycle of error accumulation that corrupts weight updates, ultimately derailing the training dynamics. To validate our findings, we introduce a minimal modification to the flash attention that mitigates the bias in rounding errors. This simple change stabilizes the training process, confirming our analysis and offering a practical solution to this persistent problem.

A situated query like "where is Lin Wei?" often encodes more than its literal content: the user may also want to know whether Lin Wei is free, in a good mood, or worth interrupting now. Standard tool-use agents answer the literal question and stop. AURA inserts an inference step between scene perception and tool use that produces an IntentFrame: a structured estimate of the implicit need with a scalar gap score that controls per-query probe budget and tool selection. On a 100-query four-scene implicit-intent benchmark, AURA improves implicit-need coverage over ReAct-style probing (Delta = +0.07, p < 10^-6); three of four scenes are individually significant, the gain reproduces on a second backbone, and a prompt ablation attributes the lift to gap calibration rather than answer memorisation. On factual lookup the controller trades raw accuracy for 82% fewer probes and zero forbidden-tool violations on a privacy-sensitive slice; scope conditions are detailed in Limitations. Code, simulator, and benchmark are released at https://github.com/innovation64/AURA.

Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter. We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks under five prompt variants each, measuring accuracy, token-probability calibration, verbal-confidence calibration, verbal parse rate, and prompt-perturbation spread for every (model x dataset x variant) cell. We find three broad results. First, evaluation design can materially change the conclusion. Switching Expected Calibration Error (ECE) token from a raw to a label-set-normalised definition changes per-cell calibration by a mean absolute 0.149. More strikingly, pairing a chain-of-thought prompt with a first-character evaluator on ARC-Challenge reduces apparent accuracy by 72-88% across all five primary models; two independent repair procedures recover 93.8% and 102.7% of the lost performance, indicating an evaluator-side rather than model-side failure. Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant. Third, prompt robustness does not track parameter count reliably. Across 10 instruct models, the correlation between model size and prompt-perturbation spread ranges from -0.244 to 0.474 across benchmarks. Taken together, these results show that reliability conclusions for small language models depend not only on the model being evaluated, but also on the evaluation pipeline used to measure it. We argue that calibration definitions, evaluator logic, verbal parseability, and prompt robustness should be reported explicitly when making reliability claims.

Evaluation budgets in agentic retrieval-augmented generation span questions, search trajectories, and repeated answers. We measure allocation precision, reading efficiency, and cost boundaries using a retrieval-feedback comparison on HotpotQA and MuSiQue. At 34.14--34.39M model tokens, broader question coverage lowers standard error by 33\% versus five reads and 12.6\% versus three trajectories. Archived nested and Q-only forecasts predict these allocations within 4.0\% and 3.5\%, respectively. Depth subsets establish no clear forecasting advantage beyond the two-trajectory audit. One-read variance penalties relative to the fitted optimum at the same token budget are 0--9.9\%, with substantial Pro uncertainty. Under recorded model fees, more questions beat more trajectories at search prices of \$0--1 per 1,000 requests; question-versus-read fee rankings remain unresolved. Temperature zero cuts answer disagreement from 14.3\% to 3.4\% while comparison precision stays similar. \par\medskip\noindentKeywords: Agentic RAG; Evaluation budget; Generalizability theory; Repeated sampling.

Ask a language model the same question about the same document twenty times, and it sometimes returns two different answers. We built Probity, a benchmark of 60 tasks and 470 items from real venture-financing filings, to measure how often this happens. Then we audited our own corpus and found a defect any excerpt-built benchmark can carry: items whose evidence is missing from the window of text the model is shown. The audit flags 36 items and separates two failures a single flag would conflate: evidence genuinely absent from the window and answers that must be computed from numbers the window does supply. Flagged items change their answers far more often, wobbling at 0.255 against 0.087 on the 427 clean items, and excluding them cuts apparent cross-model agreement by about a fifth. Before testing whether the missing evidence explains the instability, we registered a prediction: re-cut each window to hold its evidence, and instability should fall below a set threshold. It failed: the repair moved wobble by 0.058, with an interval containing zero. We report the association as correlational. Almost all measurements sit where instability cannot show, which bounds what a corpus built for accuracy can say about stability. We release the corpus, all 112,800 raw responses, and the audit as a runnable check for any document benchmark.

Motivated by recent advances in operational weather forecasting, we study the efficacy of low-precision arithmetic for climate simulations. We develop a framework to measure rounding error in a climate model which provides a stress-test for a low-precision version of the model, and we apply our method to a variety of models including the Lorenz system; a shallow water approximation for flow over a ridge; and a coarse resolution global atmospheric model with simplified parameterisations (SPEEDY). Although double precision (52 significant bits) is standard across operational climate models, in our experiments we find that single precision (23 sbits) is more than enough and that as low as half precision (10 sbits) is often sufficient. For example, SPEEDY can be run with 12 sbits across the entire code with negligible rounding error and this can be lowered to 10 sbits if very minor errors are accepted, amounting to less than 0.1 mm/6hr for the average grid-point precipitation, for example. Our test is based on the Wasserstein metric and this provides stringent non-parametric bounds on rounding error accounting for annual means as well as extreme weather events. In addition, by testing models using both round-to-nearest (RN) and stochastic rounding (SR) we find that SR can mitigate rounding error across a range of applications. Thus our results also provide evidence that SR could be relevant to next-generation climate models. While many studies have shown that low-precision arithmetic can be suitable on short-term weather forecasting timescales, our results give the first evidence that a similar low precision level can be suitable for climate.

A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt a person. Choosing well requires knowing how much the robot's sensors reveal about the cause and how reliable the robot's own diagnosis is. We build a simulated benchmark in which every failure's true cause is known, because we injected it, and measure what each sensor reveals, with explicit checks against data leakage. Some failures are diagnosable from camera images; others only from the robot's force data (0.99 from force data, no image method above 0.55). We then test six open vision-language models. Their behavior tracks the surface of the prompt, not the evidence: moving the refusal option from last to first in the answer list collapses refusal rates from 78-100% to 0-6% in three of the six swept model-and-family pairs. Accuracy from frames stays at or below a majority-class baseline under every prompt variant, with or without worked examples, and stated confidence carries no information about correctness. Handing the same models the force data as ten lines of text produces the first above-baseline diagnoses, in four of the six models: much of the failure reflects missing sensor data, not missing ability. We pose the choice as a three-action decision problem, act, consult your own sensors, or ask a human, whose optimal policy follows from measured accuracy. The models do not follow it, and their ask rates ignore a fourfold change in question cost. One question to a human still lifts them from that baseline to roughly the answerer's own reliability (0.70-0.81 when they ask). The decision to ask should be tied to measured accuracy and stated costs, not to the model's confidence.