Papers

We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task. Our key intuition is that by training the model to distinguish good and bad tokens in a response, we naturally guide the model towards generating good tokens, while avoiding the pitfalls that come with directly training the model to generate off-policy tokens. Experiments on document summarization tasks show that TOPL achieves strong out-of-distribution generalization across 11 datasets against a diverse set of sequence-level and token-level baselines. We further demonstrate that TOPL transfers effectively to machine translation, suggesting that its benefits generalize across different faithful generation tasks. Through ablation studies, we confirm that our token-level learning signal is critical to good performance; sequence-level analogues do not confer similar benefits. Finally, we show that TOPL induces interpretable model updates: the LoRA adapters learned through TOPL function as linear classification heads and steering vectors.

Recent advances in world models and video generation have given rise to an emerging reasoning paradigm that leverages video generative models to simulate, predict, and reason about real-world dynamics. We redefine this paradigm as Thinking in Video, where video is not merely an output artifact but a medium for constructing, extending, and verifying causal thought. However, this promise remains unverified: convincing rollouts may reflect memorized appearances rather than causal understanding, while existing metrics separate perceptual fidelity from semantic logic. To evaluate whether video generators support such reasoning, we introduce the Causal-Generative Dual-Judge (CGDJ), auditing World Model Consistency from two perspectives. Explicit Causal Perception tests whether a generator reads a video scenario as a reasoning problem through spatio-temporal flattened visual question answering, while Implicit Generative Perception-Prediction Gap evaluates whether it renders the causal consequence as a consistent future video. Applying CGDJ to representative open- and closed-source generators reveals a clear Perception-Prediction Gap: open-source models produce plausible dynamics despite near-zero explicit causal perception, whereas advanced closed-source systems show stronger but still limited alignment between reasoning and generation. Further analysis exposes audio-visual misalignment, where models verbalize correct causal logic more reliably than they render it, challenging the "world simulator" narrative.

Intelligence that must act before outside correction arrives must carry the structure of its expert knowledge, and hyperbolic geometry represents hierarchy with a fidelity no Euclidean space of comparable dimension matches. Yet hyperbolic language models confine that geometry to adapters or rebuild the transformer, and an independent audit of released hyperbolic vision language checkpoints finds their hierarchy mechanisms inactive, which has been read as evidence about the geometry or its objective. Here we show that it is also evidence about arithmetic. The constraint defining the manifold is a difference of two terms that grow together; for the unclamped map at unit curvature and temperature it fails in every format we measured, double precision included, at the norms standard language models produce in their deepest layers, beyond the limit previously derived. A common remedy, a norm bound, leaves the executed map no radial sensitivity where it acts. In a prespecified test, a tighter bound cost a Lorentz system more of the held out structure it started with than a flat one. Structural supervision has run in six expert systems over 20,951,552 training exposures. In paired runs with a published recipe's objective held fixed, removing its starting scale remedy prevented every update and removing its single precision regions raised the radius reached early in training. Arithmetic bounds the computable region, and within it, alongside the objective, arithmetic and its remedies select which runs leave a record and the states they reach on the way, so a flat surviving checkpoint cannot by itself establish that the geometry is inert.

Multilingual assessment systems commonly rely on translation for scoring and quality-control processes. We evaluate whether multilingual sentence embeddings can replace translated English input for Linguistic-Integrated Reliability Auditing (LiRA) across 11 PIRLS constructed-response items and three embedding models. Native-language embeddings reproduced translation-based reliability estimates closely while recovering responses excluded after translation failure, with no meaningful change in reliability.

Measuring conversational states such as cognitive load and conversational power from multimodal behavior requires characteristic features that are not only predictive but also reliable across task contexts. We present a three-dimensional evaluation framework assessing predictive accuracy, cross-task generalizability, and test-retest reliability, applied to interactional, acoustic, and linguistic features extracted from dyadic conversations during collaborative tasks performed over a video-conferencing platform (AVCAffe dataset; 53 dyads, 9 tasks). Our results show that no single feature family dominates all three dimensions. Linguistic features show the highest predictive accuracy for cognitive load but collapse under cross-task evaluation, revealing sensitivity to task-specific vocabulary. Additionally, acoustic reliability, often reported as evidence of feature stability, degrades once speaker identity is controlled, confirming that standard prosodic features measure vocal characteristics rather than conversational state. Interaction features provide the only genuinely reliable signal, unchanged after speaker normalization. Interestingly, classifying power role remained near chance baseline across all conditions, indicating limitations of task-level aggregated behavior for predicting power role in conversation. Our findings reveal three insights: (1) linguistic features predict best but generalize poorly across task contexts; (2) acoustic reliability collapses to near-zero once speaker identity is controlled, challenging standard evaluation practice; and (3) interaction features provide the only genuinely reliable signal, with floor dominance predicting within-dyad cognitive load asymmetry. These results argue for speaker normalization and multi-dimensional evaluation as prerequisites for context-aware, robust multimodal feature selection in conversational systems.

As generative artificial intelligence enters scientific and professional work, its uncertainty must be defined on the states that matter for inference and decision-making. Language models assign probabilities to words, whereas applications require uncertainty over meaningful states such as diagnoses, hypotheses or operational conditions. We introduce a semantic map: a prespecified, testable bridge from probabilities over verbal responses to a posterior over declared finite states. The language distribution remains unrestricted; held-out calibration connects it to a reference posterior. We derive posterior-error bounds and conditions for existence, conditional uniqueness, presentation stability and stable inverse recovery. This distinction matters because language probabilities depend on prompt wording, while the target posterior should not change under information-equivalent rewording. Experiments use professional market text compiled from Federal Reserve economic and financial series, together with controlled simulations having exact posteriors. Across two fitted language models, language-derived probabilities outperform printed numerical confidence, recover held-out posteriors with valid uncertainty coverage, remain largely stable under paraphrase and respond appropriately to altered evidence. Prompt engineering optimises a wording-dependent response; robust scientific use requires validated stability of application-relevant meaning. The proposed map turns semantic uncertainty in generative systems into an identifiable and testable statistical measurement problem and, when its acceptance conditions hold, yields an auditable posterior estimate.

In English all-words word sense disambiguation (WSD), the labels, not the models, have become the bottleneck: frontier LLMs are accurate enough that the errors surviving in the gold standard decide benchmark rankings – in the test sets we score on and, as we show causally, in the corpus we train on. We release lexEN, a WSD evaluation benchmark built as a conservative, human-adjudicated correction layer over Maru2022's ALL_NEW benchmark (211 labels changed, 56 removed), and SenseBench, an auditable LLM WSD evaluation harness and living leaderboard (57 models, 192 runs). The task is inventory-constrained multiple choice (the model picks from the supplied WordNet senses), so the reported accuracies are a ceiling on what models achieve without that help. On lexEN-v1 the frontier LLMs converge near 95% (best, 95.6%), the top three families are statistically indistinguishable, and accuracy trades off against reasoning effort and cost across a ~2,500x price span. Relabeling SemCor with frontier models and retraining BEM, ESCHER, and ConSeC unchanged lifts them by several F1 points on test sets the relabeling never touched; we release the relabeled corpora and Glite LENS, a 298M bi-encoder trained on the repaired labels – to our knowledge the strongest reported (83.6 Raganato ALL, 87.4 Maru ALL_NEW) – serving at ~$0.13 per million items. On hard items, fine-grained WordNet senses are partly ill-posed even for experts (three-reviewer Fleiss kappa=0.537); coarsening raises annotator agreement and model accuracy together across four inventories, placing a top model inside the expert agreement band at coarse granularity (statistically equivalent under three of four) but significantly below it at fine. The binding constraint is now cost.

Simulating societies of many large language model (LLM) agents is expensive, yet the questions asked of such simulations are usually macroscopic: phase behaviour, stylised facts, and scaling with the number of agents $N$, not the cognition of any single agent. We turn a statistical-physics observation into a method: replace each LLM agent by a low-parameter model fitted from a few hundred to a few thousand cheap queries, then run the society at any $N$ on a laptop. Whether this works is decided before the simulation runs, chiefly by what each agent perceives. We introduce an [interaction order x memory] taxonomy that maps perception and memory to an effective theory and a predicted $N$-trend of the surrogate error. We validate it on a faithful reimplementation of the LLM macroeconomy EconAgent and seven further named LLM simulations, with agent decisions cloned from genuine LLM elicitations (primarily DeepSeek) for a few dollars; the predicted error trends hold cell by cell, and the two refuted predictions, both on a strongly saturating response and traced to its curvature, are themselves matched quantitatively by the theory with no free parameters.

Traditional benchmarks for LLMs primarily rely on static datasets and objective scoring metrics, which often fail to capture differences in response quality when multiple answers are acceptable. In such settings, correctness alone is insufficient to distinguish between responses that vary in clarity, completeness, and usefulness. This paper introduces a consensus-based evaluation framework that measures relative preference among model-generated responses rather than absolute correctness. Instead of evaluating outputs against a fixed ground truth, we assess how a panel of diverse LLMs ranks anonymized candidate responses to the same prompt. This approach treats aggregate inter-model agreement as a proxy for perceived response quality under blind conditions. We conduct a controlled study using five state-of-the-art LLMs across multiple domains, including programming, general knowledge, safety, logical reasoning, and mathematics. Each model generates responses and independently ranks peer outputs through a structured voting process. Scores are aggregated into a Relative Intelligence Index (RII), representing how frequently a model's responses are preferred by other models. Our findings reveal consistent preference patterns across domains, with certain models more frequently ranked highly by their peers. However, we emphasize that these results reflect inter-model preference alignment rather than objective correctness or human judgment. This framework provides a scalable, model-driven method for comparative evaluation, offering an alternative perspective on response quality in scenarios where multiple valid answers exist. While not directly aligned with human evaluation, prior work suggests that aggregated model preferences can partially correlate with human judgments, motivating this as a proxy signal.

We present a practical recipe for building a compact Hindi text-to-speech (TTS) model by distilling a large flow-matching teacher (IndicF5, 337M-parameter DiT) under a severe data budget (~17.6 hours). Training a small model from scratch on this much data fails outright. Instead we warm-start the student from the teacher by pruning depth only: keeping the teacher's width, text dimension, attention heads, and mel/text I/O fixed so all non-block tensors copy one-to-one, and retaining an evenly-spaced subset of transformer blocks. We first measure how much depth the teacher tolerates (it remains near-functional at -27% blocks but collapses past -50%), then descend gradually (22 -> 16 -> 12 -> 8 -> 6 blocks), re-fine-tuning after each prune, with each step gated by an objective ASR word-error-rate (WER) check. The resulting students reach WER 0.00 on unseen sentences at 249M and 190M parameters, and remain robust down to 131M; at 102M we observe a clear capacity cliff that we attribute to the data budget rather than the recipe. We also document two train/inference feature- and library-parity failures (mel filterbank and rotary-embedding library versions) that silently degrade audio, and a version-independent fix. The method yields a high-quality Hindi voice that runs in real time on a 6 GB laptop GPU. An independent 50-sentence FLEURS benchmark compares the released 190M student against its teacher and MMS-TTS-hin.

Multimodal emotion recognition in conversation (MERC) can leverage multimodal and contextual cues to boost recognition performance. However, existing fusion approaches in MERC often ignore modality-specific uncertainty across utterances caused by conflicting cues, varying noise, and missing modality-specific signals. We propose EmoEUS, an explicit uncertainty supervision framework for MERC. EmoEUS performs uncertainty-aware multimodal fusion by dynamically weighting modalities using learned variance estimates. We also introduce an explicitly supervised loss that aligns each utterance's predicted variance with the distance between the utterance's distributional representation and its emotion- and modality-specific cluster center. Experiments on IEMOCAP and MELD show that EmoEUS consistently outperforms state-of-the-art methods.

Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models. We show the surgery is not clean. As a disposition probe we use 21,600 decisions under uncertainty - weekly up/down calls on 60 Warsaw Stock Exchange equities over 18 weeks, replayed through a frozen pipeline so the decision-layer model is the only variable. The task elicits no refusals at all, so any between-arm delta is pure side effect. Holding provenance constant (official BF16 checkpoints, a single abliteration author, an identical serving stack, one byte-identical frozen prompt), we compare base and abliterated arms of two Mixture-of-Experts families, Gemma-4-26B-A4B-it and Qwen3-30B-A3B-Instruct-2507. Three effects replicate across both families (weeks-clustered bootstrap CIs excluding zero): abliterated models are systematically more optimistic (+12.2 pp Gemma, +7.4 pp Qwen; the confirmed preregistered endpoint), justify themselves at greater length, and use fewer explicit uncertainty words in forced self-critiques (both exploratory). A fourth effect reverses sign: the same operation makes Gemma-abliterated less confident and Qwen-abliterated more (family CIs non-overlapping) - one weight surgery, opposite shifts in expressed confidence. Capability covariates rule out instruction-following degradation as the driver, and no arm shows economic skill: the apparent edge of abliterated arms is regime beta, not alpha. Our provenance audit also caught two independent contamination channels - a mismatched-quantizer pilot pair and a stale community chat template that silently mangled the rendered prompt - suggesting toolchain artifacts are the rule in studies of community-modified checkpoints. Whoever deploys an "uncensored" model as an agent is deploying a measurably different decision-maker, not the base model minus refusals.

Biographical personas are widely used in system prompts, but their effects on code generation are rarely evaluated under controlled, pre-registered conditions. We tested four prompt conditions (no persona, two engineer personas, and a research-librarian persona), 12 code-generation tasks, two frontier models, and five runs per cell (480 completions). Persona effects differed between the two tested models. Under the pre-registered mixed-effects analysis, the condition-by-model interaction was significant for provider-reported output tokens; a post-hoc visible-character measure showed the same qualitative pattern. Six GPT-5.5 completions were length-capped and are reported separately. On Claude Opus, the minimalist engineer persona reduced visible output by 30% (33% in provider tokens) without improving correctness, while the thorough engineer persona increased output without a correctness gain. In an exploratory post-hoc analysis, the librarian persona elicited in-character disclaimers in 55 of 60 Opus responses and 12 genuine no-code responses, lowering mean correctness from 0.92 to 0.67. GPT-5.5 produced neither behavior in its 59 non-truncated responses. These results are consistent with personas acting as Model-Dependent behavioral-policy biases rather than universal quality interventions. We release raw completions, derived scores, analysis artifacts, a pre-registration document, and an execution gate log; end-to-end test-based rescoring requires an unreleased task harness.

Existing research shows that AI-generated text detection classifiers achieve strong in-distribution (ID) performance but do not maintain the same performance on out-of-distribution (OOD) texts, suggesting overfitting to dataset-specific features. However, combining different training datasets doesn't always improve performance and, in some cases, can even encourage shortcut learning. To address this issue, we fine-tune BERT-tiny models with Bayesian classification heads to select texts across three different datasets to use as a consolidated training set. We trained three different classifiers: fine-tuned DeBERTa-V3-large and ModernBERT-large classifiers via empirical X-risk minimization, and an MCGrad model that calibrates the predictions from the ModernBERT-large classifier. The DeBERTa-V3-large-large classifier achieves a mean score of 0.882 on the PAN 2026 test set across five metrics: AUROC, $F_1$, C@1, Brier score, and $F_{0.5u}$. ModernBERT-large achieves a score of 0.96 while MCGrad achieves the best score of the three with a mean score of 0.974, ranking second on the leaderboard. Our results highlight that careful dataset curation can lead to strong OOD performance. We release our ModernBERT-large and DeBERTa-V3-large models at https://huggingface.co/collections/ShantanuT01/panclef-2026 .

Multimodal emotion recognition in conversation (MERC) achieves accurate predictions by integrating multimodal and contextual information in dialogues. While current MERC approaches focus on modeling complex contextual dependencies in conversation, they often overlook the impact of contextual emotional inertia in emotion shift, leading to sub-optimal performance. To address this issue, we propose a novel Emotional Inertia-Informed Supervised Contrastive Learning module (EII-SCL) that informs the contrastive objective by constructing inertia-affected samples within temporal windows, effectively leveraging emotional inertia as a prior while enabling seamless integration with existing MERC models without requiring additional data. Extensive experiments on IEMOCAP and MELD show that our approach consistently outperforms state-of-the-art methods.

Long-horizon rollout generation has become the dominant systems bottleneck in agentic reinforcement learning (RL). As agents interact with environments over many turns, trajectories rapidly grow to tens of thousands of tokens, making synchronous RL training increasingly constrained by rollout. We propose WAR, a workload-aware rollout system that substantially accelerates synchronous agentic RL by jointly optimizing decoding and scheduling. WAR is built on a key observation: the optimal rollout optimization strategy depends on runtime load: (1) Under low load, WAR enables model-free speculative decoding with SuffixDecoding, which reuses suffix patterns from previously completed trajectories as speculative drafts for future rollouts. Unlike model-based drafters, SuffixDecoding introduces no additional draft model and avoids GPU contention with rollout generation. (2) Under high load, where saturated batched decoding leaves limited room for speculative speedup, WAR shifts the optimization focus to cache-aware scheduling. A global scheduler places requests across rollout replicas based on cache locality, trajectory progress and server load, reducing redundant KV-cache recomputation and mitigating load imbalance. By combining decoding-level suffix reuse with system-level rollout scheduling, WAR delivers robust throughput improvements across workload regimes without changing the underlying RL algorithm. WAR improves long-context agentic rollout throughput by 1.4x under low load and up to 1.6x under high load. These results show that WAR removes a major rollout bottleneck in synchronous agentic RL and provides a practical path toward scalable long-context agent training.

Deep research agents increasingly operate over the open web, where relevant records coexist with redundant summaries, outdated reports, and misleading documents. Existing evaluations offer limited insight into whether agents preserve sound evidential standards when an ordinary-looking false document is deliberately seeded into a searchable environment and offers a direct shortcut to a conflicting answer. We introduce DRNOISE, a 100-task benchmark for answer recovery under misleading evidence. Each task has a unique gold answer supported by two corroborating indirect record chains; the paired noisy condition adds one plausible document that states a conflicting answer directly. The benchmark spans ten families of evidence operations. Among models with strong clean-task performance in the shared agent framework, this single intervention causes 66–88 percentage-point accuracy drops. Trace analyses reveal a pattern of verification inertia: agents often retrieve truthful records yet defer to the answer-like document without adequately checking its claim against the underlying evidence. Generic verification prompts reduce but do not close this gap. The setting is especially relevant to open-web deployment, where plausible falsehoods arrive through ordinary-looking pages without explicit adversarial instructions. Reliable deep research therefore requires more than retrieval and citation; it requires active reconciliation of direct claims with record-level evidence.

Safety evaluation of large language models is conducted predominantly in English and predominantly on frontier systems. Neither condition describes how such models are encountered in low-resource health settings, where small quantised systems are run locally and queried in local languages. We ask whether clinical safety established in English transfers to Hausa, and whether any failure is attributable to the language, the clinical task, or the class of model that low-resource deployment admits. Matched English-Hausa question pairs were built for three conditions of high burden in northern Nigeria: malaria, sickle cell disease, and tuberculosis, probing knowledge recall, emergency triage, a leading question inviting a contraindicated action, and a traditional-remedy claim. Six models were evaluated: five locally deployable systems of 4-9 billion parameters, two medically fine-tuned, and one frontier system. All 128 responses were scored against Nigerian national treatment guidelines by two fluent Hausa speakers working independently and blind to one another. Among locally deployable models, mean clinical correctness fell from 1.57 in English to -0.03 in Hausa, on a scale where 2 denotes a correct answer and -1 an actively harmful one. The frontier model moved from 2.00 to 1.75 and produced no response judged harmful in either language. Drift was consistent across all three conditions. Inter-rater agreement was substantial for clinical correctness (kappa = 0.70); agreement on harm was initially poor (kappa = 0.22) and is examined in detail. Because a frontier model answers the same questions competently in Hausa, the deficit is a property neither of the language nor of the clinical material, but of the deployable tier.

Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing. However, LLMs often suffer from hallucinations and lack of relevant knowledge when dealing with question answering (QA) tasks. To mitigate these issues, knowledge graphs (KGs) have been utilized to enhance LLM reasoning. Nevertheless, KGs often contain noise and errors, while existing KG-enhanced LLM approaches are generally unable to identify and filter such noisy and erroneous content, which can instead amplify hallucinations and pose challenges for reliable reasoning. Uncertain knowledge graphs (UKGs), which associate each triple with a confidence score to quantify uncertainty, offer a promising direction to address this challenge. Compared with prior work, we investigate how to leverage UKGs to support LLMs for QA. We propose Debate-on-Graph (DoG), a new framework that enables LLMs and UKGs to collaborate adaptively for reliable reasoning. Specifically, we first design a heuristic search algorithm tailored for UKGs to extract reliable and question-relevant subgraphs, thereby reducing noise and errors in retrieved knowledge. We then introduce a Multi-Agent Debate mechanism, which yields reliable answers through adaptive adversarial debates, aiming to fully exploit the knowledge in UKGs while preserving the reliability of retrieved evidence. Extensive experiments on four benchmark QA datasets show that DoG achieves state-of-the-art performance over existing LLM reasoning methods and KG-based baselines, while enabling reliable and adaptive reasoning. Our code is available at https://github.com/seucoin/Debate-on-Graph.

Existing methods for multimodal sentiment analysis (MSA) under missing modalities usually follow a repair-first paradigm. We revisit this assumption and ask: should every missing modality be repaired? A per-sample oracle analysis shows the answer is not always: full-modality input is optimal for only a small fraction of samples, and every modality subset is preferred by some samples. These results suggest that adding or repairing modalities may not always improve prediction, and that the utility of each modality is sample-dependent. Building on this finding, we propose Sufficiency-Informed Evidential ValvE (SIEVE) that turns “whether to repair” into an explicit, learnable decision at the sample level. SIEVE compares a direct prediction branch with a repair branch, derives an empirical sufficiency signal from their per-sample loss gap, and routes each input through an evidential gate that jointly models sufficiency and its epistemic uncertainty. SIEVE is repair-agnostic: it operates as a plug-and-play decision on top of any explicit or implicit repair module, without modifying its internal design. Experiments on CMU-MOSI and IEMOCAP show that SIEVE consistently improves representative repair backbones across evaluated missing rates, and approaches the per-sample dual-branch achievable optimum.

This paper introduces EvolvingWorld, a framework and benchmark for character and world co-evolution in interactive literary worlds. Existing systems either treat interactive literary simulation as static persona imitation or isolated scene generation, failing to capture how characters and worlds evolve together over time. To address this, EvolvingWorld models literary simulation as a long-horizon process where characters interact, scenes progress, and character and world states are persistently updated. Unlike prior systems relying on fixed schemas, EvolvingWorld adopts an open-schema framework to support simulation across diverse literary worlds. The framework consists of two coupled modules: a Character Agent for multi-character role-play and persistent profile evolution, and an LLM-based World Model for global and location/entity-level state maintenance and scene progression. Based on this architecture, we formulate 7 trainable tasks for scene initialization, interaction generation, and state update. We construct a dataset from 57 books, producing 138,596 supervised training samples and 222 snapshots for testing. Furthermore, we introduce a trajectory-level LLM-as-Judge evaluation protocol spanning 10 dimensions and 20 metrics. Experiments show that EvolvingWorld can improve long-horizon simulation by effectively maintaining persistent, coherent character and world development.

AI_LectureNote is a historical, readability-oriented post-ASR workflow for Korean-English medical lectures. It rewrites speech-to-text output into study transcripts while restoring Latin-script medical terms rather than Korean phonetic transliterations. We retrospectively evaluate the workflow on four author-recorded lectures across five conditions. In this pilot, post-processing raised the macro English-script rendering rate from 0.39 to 0.71 on the whisper-1 path and from 0.26 to 0.65 when applied to 3-minute chunked gpt-4o-transcribe output. However, English-script rendering did not imply semantic faithfulness: the two post-processed conditions showed semantic drift in 34 and 36 of 282 reference sentences and polarity failures in 11 and 13 of 101 polarity-cue rows. A descriptive cross-input comparison suggested different candidate failure patterns: polarity-failure sets overlapped more strongly across front-ends (Jaccard 0.60; 9 shared of 15 unioned failures) than general semantic-drift sets (Jaccard 0.23; 13 shared of 57 unioned drifts). This single-annotator pilot documents concrete failure modes rather than population rates and supports evaluating surface accuracy, term-script rendering, chunk-level script consistency, and medical-meaning preservation separately.

This paper describes TalTech's submissions to the Beyond Transcription Challenge (BeTraC), which requires generating SOAP notes directly from long doctor-patient conversation recordings, without intermediate transcription. After screening open-weight speech LLMs for long-audio robustness, we adapted Voxtral Mini (lightweight track) and Voxtral Small (heavyweight track) with LoRA supervised fine-tuning followed by DAPO reinforcement learning that uses the challenge metric, Open Medical Concept F1, as its reward. Our systems ranked first in both tracks, and an independent LLM-as-a-judge evaluation showed the lowest hallucination rate among all submissions, indicating that reinforcement learning against a concept-matching metric need not compromise factual reliability. We also find that fine-tuning on text transcripts transfers well to speech input and appears to improve robustness on out-of-domain real recordings.

Prior work on LLM-generated text has demonstrated quantitative and qualitative departures from text produced by humans. LLM-generated texts differ from human writing in style, resulting in a characteristic textual "feel," while the semantic range of LLMs is much restricted compared to that of humans. In this contribution, I note simple but consistent patterns in the statistical distribution of n-grams within LLM-generated text. Via qualitative analysis of these n-grams, I reveal deficiencies in LLM style. Because higher-order n-grams correlate to semantic content, I conclude that questions of style and semantics are not cleanly separable.

Question-order effects in human survey data have been reported to approximately satisfy the QQ (quantum question) equality, a parameter-free prediction of the standard projective quantum question-order model. We develop this equality into an audit framework for sequential binary judgments of autoregressive large language models (LLMs). Theoretically, we characterize mechanism families that satisfy QQ robustly, show that classical repetition can reproduce the equality exactly, and combine QQ with the rank-2 Contextuality-by-Default criterion through $|q_{QQ}| \le \mathrm{OSS}$. This separates order sensitivity, QQ imbalance, and residual contextuality rather than treating them as interchangeable signatures. Methodologically, we introduce a committed multi-turn forced-branch protocol that reconstructs order-conditioned joint distributions from next-token log-probabilities under counterbalanced label mappings and pre-specified health gates. A first-signal pilot on an open-weight instruction-tuned model reveals the central measurement problem. Although all pre-specified health gates passed, the binary-conditioned distributions were near-deterministic for 17 of 18 item pairs under the direct-evaluation framing and 7 of 8 under the persona framing. Label assignment materially changed several mapping-specific QQ verdicts, and no item was certified as residually contextual. Thus, under the tested conditions, the observed QQ outcomes did not uniquely identify a response mechanism in the presence of a saturated and label-sensitive measurement interface. The main implication is methodological: next-token probabilities should not be interpreted as survey-response distributions without first establishing adequate dispersion. We therefore argue that saturation screening and label counterbalancing should precede structural interpretation in distribution-level audits of LLM judgments.

Evaluating large language models (LLMs) across languages remains challenging, as most multilingual benchmarks rely on translated English datasets, often obscuring linguistic and cultural specificity in the target language. This issue is particularly pronounced for less-resourced languages such as Kyrgyz, where reliable natively authored evaluation data are scarce. Building on previously introduced Kyrgyz-language evaluation datasets, this work reports the first systematic and large-scale evaluation of LLMs in Kyrgyz using the KyrgyzLLM-Bench benchmark suite. KyrgyzLLM-Bench comprises two natively authored datasets$-$KyrgyzMMLU and KyrgyzRC$-$together with carefully translated and manually post-edited versions of WinoGrande, HellaSwag, BoolQ, and TruthfulQA. We evaluate 26 open- and closed-source LLMs under zero-shot and few-shot settings, analyzing model performance, cross-lingual transfer, and the impact of translation artifacts on evaluation reliability. Across families and tasks, model rankings transfer broadly from English to Kyrgyz on WinoGrande and BoolQ, and to a lesser extent on MMLU, while HellaSwag exhibits a substantial English-Kyrgyz performance gap consistent with translation-induced plausibility shifts. Few-shot prompting improves several open-source models on reading comprehension but behaves inconsistently for proprietary models on translated tasks. We publicly release all datasets, evaluation code, and per-model results, and integrate the Kyrgyz tasks into a widely used multilingual evaluation framework to support future research on Kyrgyz NLP.

Jailbreak attacks on large language models are usually evaluated by attacker-centric metrics such as attack success rate (ASR), yet an attack that breaks a model is not necessarily useful for improving its safety. We propose a defender-centric view of jailbreak evaluation, where attacks are evaluated by the downstream safety improvements they enable when used as red-teaming data for safety training. Building on this view, we introduce A-MESS (Minimal Effective Attack-Subset Selection), a setting-agnostic framework for attributing and selecting jailbreak attacks from black-box subset utility observations. A-MESS estimates AttackSHAP, a Shapley-based score that attributes marginal utility to individual attacks and selects compact attack subsets under user-specified budgets via greedy or surrogate-based optimization. Across controlled utility landscapes and real LLM safety settings, we find that ASR rankings are weakly aligned with defender-centric utility, that AttackSHAP can be estimated accurately with limited utility queries, and that directly optimizing subsets yields stronger safety utility than attacker-centric or attribution-only selection. These results suggest evaluating jailbreak attacks as resources for improving safety, not only as tools for breaking models.

LLM agents can leak privacy (e.g., paths, emails) and credentials (e.g., API keys) as agent observations (e.g., tool outputs, shell logs, and file reads) are appended to provider-bound transcripts. Existing placeholder redaction is brittle: it can miss embedded or cross-turn references, over-redact benign lookalikes, and destroy the structure useful for reasoning. We present SlotGuard, a local transcript boundary that can hide sensitive data while retaining agents' performance. SlotGuard rewrites structural bindings as typed, suffix-aware slots, replaces secrets with format-preserving synthetic values, links cross-turn references with a lightweight session graph, and restores raw values only inside the trusted runtime. On controlled repository-oriented agent transcripts, SlotGuard removes all 20,814 annotated structurally sensitive characters across 9,229 paths and reduces credential leakage to 0.0% across 852 planted values. It remains close to raw-transcript task success across four upstream models, while generic redaction drops to 2.5%. Transcript rewriting takes a median of 14.424~$\mu$s per agent turn. The code is publicly accessible at https://github.com/illinoisdata/SlotGuard.

We present a continuous geometric framework that models the discrete algebraic operations of the Transformer architecture as an integro-differential equation (IDE) on a semantic fiber bundle $\calE = \calM \times \R^d$. Beginning from a single geometric axiom – that the token sequence forms a discrete $1$-manifold equipped with a canonical measure lattice – we translate every core component of the modern Transformer (RMSNorm, RoPE, Softmax Attention, FFN, Residual Stream, SGD, Weight Decay) into a cohesive vocabulary of differential geometry, measure theory, and stochastic calculus. The resulting framework yields quantitative predictions spanning entropic optimal transport (Attention as a Schrödinger bridge) and non-equilibrium thermodynamics (SGD as Itô diffusion violating detailed balance). We conduct a six-part experimental campaign across five architectures (Qwen3, LLaMA\nobreakdash-3.1, Gemma\nobreakdash-3, GPT-2, Mistral) spanning $124$M to $8$B parameters. The empirical observables are quantitatively consistent with the geometric predictions: the $\epsilon^{-1/2}$ Lipschitz scaling calibration at machine precision ($R^2 = 1.000$), the Lie–Trotter operator-splitting torsion, the symmetric ablation instability confirming the Dual-Law of Topological Stability, the $\calO(1/\sqrt{k})$ thermodynamic suppression of Poincaré recurrence on the RoPE torus, the thermodynamic context-limit phase transition, and the Non-Equilibrium Steady State parameter vortex – verified across two optimizers (AdamW and Pure SGD) to exclude momentum artifacts. The results demonstrate that analyzing Transformers through the lens of continuous stochastic differential geometry provides a predictive descriptive vocabulary for the stability limits, context bounds, and optimization dynamics of Large Language Models.

Prior case retrieval (PCR) aims to identify the precedent cases relevant to the facts of a query case. Incoming citation context, the text with which later cases characterize a case when citing it, is a powerful relevance signal, yet it is typically evaluated without a temporal constraint, so the retriever is credited with citations made after the query. We introduce a temporally fenced retriever with no learned parameters that augments BM25 with incoming citation context restricted to citations predating the query, together with a temporal-admission decomposition that quantifies the phantom fraction: the share of a citation-context gain attributable to citations not known to predate the query. Experiments span two jurisdictions, U.S. federal (CLERC) and European (ECtHR-PCR) case law. On ECtHR-PCR, without any training, the fenced retriever outperforms a strong degree-controlled baseline across the full recall ladder, and a temporal-admission decomposition attributes 14.9% (validation) of an unfenced citation-context gain over BM25 to citations not known to predate the query. Citation-context retrieval must therefore be temporally fenced and degree-controlled before its reported gains can be interpreted.