Papers

Moral language plays a central role in shaping online endorsement and the diffusion of information, yet existing moral foundation detection systems often suffer from poor cross-domain generalization, weak rationale grounding, and reliance on costly prompting-based large language models (LLMs). We introduce CHARM, a MAC- and Hate-speech-Aware Rationalealigned Moral foundation detection framework built on a lightweight fine-tuned LLM, which integrates complementary moral grounding, rationale alignment, and polarity-aware hate speech signals to support more robust and faithful moral prediction. Unlike prior dictionary-, fine-tune-, or prompt-based detectors, which decouple computation from psychological theory, CHARM is built so that each component – MAC cross-attention, rationale alignment, and hate-speech modulation – operationalizes a distinct psychological construct. Using a 30% subsample of the MFTC, MFRC, and News training pools together with the richer supervision in MFTCXplain, CHARM improves AUC by up to 15.3% in-domain, surpasses the supervised baselines on every out-of-domain dataset in both AUC and F1, and offers a scalable, low-cost alternative to prompting-based LLM detectors. We further apply CHARM to large-scale COVID-19 discourse on Twitter and show that moral value alignment is strongly associated with online endorsement behavior. By making moral framing measurable at scale, CHARM offers a practical tool for studying the spread of morally charged misinformation. Code and additional materials: https://github.com/HuixiangF/CHARM/.

Language models encounter typos, corrupted text, altered words, and disrupted token order, yet robustness is usually evaluated only through output behavior. We study how six naturalistic and synthetic input perturbations propagate through decoder-only language models at three levels: output behavior, hidden-state geometry, and attention-head function. We evaluate behavioral effects across four GPT-2 and two Qwen2.5 checkpoints, analyze layerwise geometry using centered kernel alignment and intrinsic dimension, and examine attention-head responses in GPT-2. Perturbation types produce distinguishable metric profiles that are not fully captured by output measures and are only partly consistent across the tested checkpoints. Copying scores show the strongest pooled associations with activation-patching recovery under token substitution and shuffling, although these associations do not isolate copying-specific effects. Gradient-guided HotFlip perturbations also cause stronger behavioral and representational disruption than rate-matched random token substitutions in GPT-2; their behavioral effects are consistent across all six tested checkpoints. Our results show that robustness claims based on a single behavioral or representational metric can be misleading, and motivate multi-level evaluation of how perturbations alter language-model computation.

The Neural Finite State Machine (NFSM) framework offers a pragmatic path to full-duplex dialogue by serializing turn-taking control and response generation onto a single causal tape under the standard next-token prediction objective, thereby preserving semantic prowess at a low fine-tuning cost. However, its reliance on synthetic text data fundamentally limits turn-taking naturalness, as Large Language Models (LLMs) cannot faithfully simulate the fine-grained acoustic temporal dynamics of real human dialogues. In this work, we propose a decoupled data approach that learns turn-taking from real Human-Human (HH) spoken dialogues while shaping semantic behavior through configurable Human-Agent (HA) text dialogues. To operationalize this approach, we introduce a rule-based event-guided data transformation method that serializes HH spoken dialogues into FSM tapes by classifying turn-taking events and applying deterministic mapping rules, enabling scalable supervision without LLM-generated annotations. We further propose a Source-Aware Calibrated (SAC) Loss that jointly calibrates the long-tailed distribution of state transition tokens and channels each data source toward the capability it best supervises. Experiments show that our approach substantially improves turn-taking proficiency while recovering the foundation LLM's semantic capability. Our code and model are available at https://github.com/Liyht/def-fsm.

Personalized assistants should not only comply with user requests but also assess whether those requests are appropriate given the user's current circumstances. However, prior work has primarily focused on accurately executing requests, overlooking the need for assistants to account for context and engage in conflict-based refusal. Furthermore, while existing work on conflict or safety detection relies on explicitly provided factors, real-world scenarios often involve implicit factors that must be retrieved from a knowledge base (KB). To this end, we introduce Personalized Assistants for Conflict Evaluation (PACE), a dataset for evaluating whether models can identify latent constraints, expressed as egocentric knowledge or events, that render seemingly reasonable user requests inappropriate. PACE pairs user requests grounded in well-defined personas with egocentric KB facts, requiring models to integrate contextual evidence to determine whether a request is conflicting. This implicit retrieval setting hinders the direct association between user requests and conflict-inducing knowledge, making it difficult for existing models to identify relevant user-specific facts. To address this challenge, we further propose PaceMaker, a multi-agent framework in which specialized agents coordinate across query reformulation, multi-hop graph traversal, and conflict-aware filtering to retrieve contextually decisive evidence. Experiments on PACE evaluate both evidence retrieval quality and conflict decision accuracy, showing that PaceMaker consistently outperforms existing approaches.

Tamil spell and grammar correction is challenging because Tamil is an agglutinative low-resource language with rich verbal morphology, complex sandhi (phonetic transformation) rules at word boundaries, and a script of 247 distinct letters. Prior work targets word-level surface errors with rule-based methods, statistical n-gram models, Minimum Edit Distance, or hybrid pipelines with a transformer re-ranker; such methods cannot reliably handle contextual errors - subject-verb agreement, tense consistency, or cross-word sandhi - which require sentence-level understanding. We propose an end-to-end sequence-to-sequence formulation and fine-tune mT5-small and mBART-50 on a synthetic corpus of up to 657,720 noisy-clean Tamil sentence pairs spanning ten error categories. Both backbones follow the same four-stage progressive schedule, each stage targeting one weakness: surface noise (v2), contextual grammar (v3), single-site sandhi (v4), and multi-site cross-word sandhi (v5). On a 1,000-sentence balanced diagnostic set verified disjoint from all training data, our best model, mBART-50 v5, reaches 69.3% top-1 exact-match accuracy, with 87.5% on sandhi and 43.5% on subject-verb agreement. The schedule is what produces these gains: subject-verb accuracy rises from 1.0% to 52.5% once contextual pairs are introduced, and sandhi from 0% to 87.5% once multi-site sandhi pairs are. We additionally quantify a precision-recall trade-off this literature has not reported: sandhi recall is paid for monotonically in identity accuracy. Finally, Tamil-LLaMA-7B-Instruct reaches 19.0% zero-shot and 24.7% with three demonstrations against a 20.0% copy baseline, showing that a Tamil-adapted instruction model does not transfer to specialised sentence-level correction without task-specific supervision.

A benchmark score credits final answers, but not the route by which an item can be answered. In medical multimodal multiple-choice questions (MCQs), this distinction matters because a correct answer can be supported by the intended image finding or by benchmark-preserved cues in the wording of answers, non-visual clinical text, visible image text, artificial annotations, or device/context artifacts. We call the resulting score-level overinterpretation reasoning inflation. Here, a route is an observable input path that can support answer selection, not a claim about the model's hidden cognition. Across six medical multimodal MCQ datasets, we separate candidate cues from behavioral evidence through prompt- and image-side audits, modality ablations, and matched repairs that preserve the medical target and answer key. In a 13-configuration open-model panel, full-input accuracy is 62.63%, while text-only and options-only settings achieve 53.96% and 29.71%, respectively. Removing length-gap, absolute/conspicuous, and spatial/prepositional cues lowers accuracy by 6.58, 3.50, and 4.77 percentage points. We also construct MedQA-MM, a 1,000-item shortcut-mitigated subset, where text-only and options-only accuracy fall to 5.21% and 12.33%. This does not imply that models never use images; it shows that medical image-reasoning claims require route-level evidence.

Large Language Models (LLMs) often help users generate artifacts through iterative cycles of generation and revision in conversation. A challenge here is that, when users specify only a local change during revision, LLMs must instead identify the relevant dependencies and propagate the revision to all affected parts of the artifact. This paper studies this ability of LLMs on conversationally generated artifacts, where the artifact context and its dependencies may be embedded in the conversation history. Toward practical use, we also explore cost-effective test-time compute for this new setting. Specifically, we introduce a new benchmark for this setting, and evaluate nine revision methods, including sequential reflection and parallel sampling variants, using gpt-oss-20b/120b, gpt-5.4-mini, and qwen3.5-9b/27b/122b on the benchmark. The results show that baselines achieve accuracies of 68.3–93%, and the most cost-effective method is selecting from three parallel samples using either LLM-based or medoid selection, which improves accuracy by 2.2–9.7%. Our code and dataset are available at https://github.com/ntt-dkiku/llm-revision-propagation.

Large language models (LLMs) face severe memory bottlenecks in long-context inference due to the linearly growing size of key-value (KV) caches. Existing KV cache compression techniques typically rely on simple heuristics, overlooking the distinct functional roles of different attention heads. We present SGD-KV (Summarization-Guided KV Cache Compression), a head-aware framework that leverages a novel chunk-summarization diagnostic task to systematically identify and prioritize attention heads specialized in hierarchical information aggregation. Experiments on Qwen2.5-7B-1M and Qwen3-32B across diverse long-context benchmarks demonstrate that SGD-KV achieves state-of-the-art performance with contexts up to 1M tokens, while reducing KV cache memory usage by up to 75%. Our findings show that strategically allocating the KV cache budget based on the summarization score distribution of attention heads yields a superior efficiency-accuracy trade-off for long-context inference.

Predicting student performance from educational interaction data requires models that are both accurate and sufficiently transparent to support meaningful intervention, while demographic information introduces an additional risk of unfair predictions. This study investigates a multi-objective framework that combines reinforcement learning-based multiple instance learning (RL-MIL), adversarial debiasing, and preference-conditioned hypernetworks for student-at-risk prediction. MIL represents each student as a bag of weakly labeled interactions, while an RL agent selects informative instances for downstream classification. Two hypernetwork variants are evaluated to determine whether a user-defined preference scalar can continuously control the trade-off between predictive performance and Equalized Odds. The underlying RL-MIL baseline achieves strong classification performance, but both hypernetwork extensions exhibit mode collapse: changing the preference weight produces little systematic movement along the intended fairness-performance frontier. The failure is associated with objective dominance, weak gradient propagation through the conditioning mechanism, and interactions between dynamically generated parameters. The results show that fairness objectives can be incorporated into an interpretable RL-MIL pipeline, but preference conditioning alone does not guarantee controllable multi-objective behavior. Robust fair RL-MIL therefore requires explicit mechanisms for gradient balancing, objective separation, and stability analysis.

Private domain speech is difficult to collect and redistribute, while compact models need task-specific supervision. We study an auditable synthetic pipeline that maps Japanese care handoffs directly to six-field structured notes. Using 182 synthetic training and development clips, we adapt a 1.47B audio model by full fine-tuning and rank-16 LoRA. On a 39-clip scenario-seed-disjoint synthetic test, an unadapted model obtains a model-judged factuality-recall score of 0.0500, full tuning 0.8664, and LoRA 0.8461. LoRA reaches 97.7% of the full-tuning aggregate as a descriptive ratio while project telemetry reports 12.4M trainable parameters, about 0.85% of the backbone. Both adaptations show large paired gains over the same base; the full-versus-LoRA interval crosses zero, and differing optimization settings preclude an equivalence claim. This is a parameter-efficient capability-acquisition result, not a device-performance result: latency, memory, energy, and real-time factor were not measured. All evaluation speech and targets are synthetic, references are model-proposed, and the judge is uncalibrated. The evidence shows that a compact model can acquire a narrow audio-to-structure transformation from a few hundred provenance-linked synthetic examples; it does not establish clinical validity, real-speech transfer, or superiority to a clean cloud system.

Language models used in financial backtests suffer from look-ahead bias, as a model trained on text published after the study period has already observed the outcomes it is asked to predict. To handle this issue, point-in-time (PIT) language models are pretrained on chronologically filtered corpora and released as one checkpoint per calendar year, each with a documented cutoff. However, each additional year costs another pretraining run, and whether that run is necessary has never been tested. In this paper, we show that a newer checkpoint is not detectably better on the decision it informs, as its information coefficient on a fixed evaluation window shows no decay with age that we can detect. Motivated by this observation, we propose PALM (Point-in-time Adaptation for financial Language Models), a simple yet effective alternative to annual pretraining that fits a low-rank adapter on financial text the checkpoint is allowed to read, without modifying any pretrained weight. We further find that the gain comes from the financial domain rather than from recency, which leaves the point-in-time constraint without a detectable cost. We validate PALM on a decade of financial news and on three families of PIT models, whose cutoffs span two decades. The pooled gain keeps its sign across experimental settings and comes at a small fraction of the cost of pretraining. Code is available at: https://github.com/seunghan96/palm.

Peer review plays a central role in scholarly publishing, yet verifying whether reviewer claims are supported by manuscript evidence remains a largely manual and time-consuming process. We present Peerify, a pipeline for manuscript-grounded verification of peer-review claims. Given a manuscript and a review comment, the Peerify pipeline decomposes reviews into atomic claims, retrieves relevant manuscript evidence, and determines whether each claim is supported by the paper. To support the development and evaluation of the pipeline, we construct a benchmark of 800 claims derived from authentic peer-review interactions collected from NeurIPS 2024 and ICLR 2024, including a 300-claim hand-labeled subset used to audit the automated supervision. We evaluate state-of-the-art language models and retrieval strategies within the Peerify pipeline, together with entailment baselines. Our results demonstrate the importance of retrieval-centered verification and claim decomposition, while highlighting the challenges posed by ambiguous and interpretive reviewer claims. Automated labels agree with human consensus on 90.3% of audited claims ($\kappa = 0.87$), while off-the-shelf entailment models stay below 0.24 macro-F1.

Large language models (LLMs) exhibit a range of human-like decision-making behaviors, but whether these reflect similar underlying mechanisms or surface-level mimicry remains unclear. We evaluate whether LLM context sensitivity aligns with a cognitive economic theory that explains human behavior through problem categorization and attention allocation. Across 12 open-source and commercial LLMs on a novel 140,000-trial product choice benchmark, context induces human-like shifts in choice and problem categorization, but does not reliably reweight attention between features like price and quality. Neither scale nor chain-of-thought reasoning reliably attenuates context sensitivity or generates human-like behavior. These results suggest that LLM decision mechanisms are distinct from human ones.

A direction in activation space that changes safety-relevant behavior when steered is not necessarily one the model uses to produce that behavior on its own. We therefore ask whether steering acts through the computation of the unmodified model or through a different set of components, studying two traits whose directions have been extracted and validated in prior work: refusal and sycophancy in Qwen2.5-7B-Instruct. For each, we use the trait vector to split the computation into a reconstruction circuit before the vector and a transmission circuit after it. We then test whether restoring the coordinate returns behavior removed by ablation. For refusal, both circuits are compact and faithful, restoring the coordinate alone recovers almost all of the refusal signal lost to ablation, and a circuit built around the vector matches the faithfulness of a direct input-to-output circuit at roughly half the edges. For sycophancy, transmission is compact, but reconstruction is broader and only partially faithful. Circuits that transmit a steering intervention are therefore not automatically the circuits that produce the behavior, and we report the counterfactual, target, and behavioral test for every circuit accordingly.

Long-form factuality verification is commonly implemented as a static decompose-search-verify pipeline, with separately prompted modules processing claims and invoking external search. Treating claims independently makes LLM and search calls scale with claim count and causes repeated searches for overlapping evidence about related claims. We introduce EAVer, an End-to-end Agentic Verifier that learns to control the complete response-level verification workflow as a unified policy. EAVer groups semantically related claims, routes each group to direct verification or targeted search based on confidence, and keeps evidence returned by search in compact in-context memos for cross-claim reuse. To train this policy, we develop a privileged-teacher synthesis pipeline that converts gold claim annotations into executable multi-turn tool-interaction trajectories with live search rather than post-hoc rationales. Structural, label-alignment, tool-use, search-budget, and leakage checks yield 1,447 quality-controlled trajectories. We further construct 794 bidirectional same-trajectory preference pairs that keep claim grouping, search, and evidence fixed, enabling decision-focused Direct Preference Optimization (DPO) over factuality-decision tokens. The results with Qwen3-8B show that EAVer outperforms the strongest search-based baseline on each benchmark by 2.88 Macro-F1 points on VeriFastScore and 4.73 points on the out-of-distribution FaStFact-Bench, while using about 80% fewer searches than the most search-efficient baseline. Moreover, EAVer consistently improves performance across models ranging from 4B to 32B parameters, demonstrating its strong generalizability.

Large language models can generate executable data-analysis code, but successful execution is not equivalent to a valid official-statistics result. This study asks whether authoritative metadata and execution feedback improve the reproducibility of Eurostat answers produced by a coding agent, and isolates what execution feedback actually contributes. A benchmark of 30 natural-language tasks covering seven domains, seven Eurostat datasets and four difficulty tiers was run under four conditions: task only (A), task plus a frozen dataset metadata card (B), metadata plus a repair loop driven by sanitized execution feedback (C), and metadata plus the same attempt budget with no diagnostics of any kind (D). Claude Sonnet 5 generated Python through the Anthropic Messages API in three independent replicates, yielding 360 task-runs. Exact correctness required successful execution, the correct dataset, filters, output shape, values and unit. A companion experiment run under an under-specified output contract, in which the required ranking key and unit representation were never stated to the model, understated condition C by 23.4 points, showing that evaluator and contract design can dominate measured agent error. Reliable statistical coding agents need semantic validation against frozen specifications, a fully specified output contract, and a retry budget - not execution diagnostics.

Large language models (LLMs) can generate fluent and coherent text that is increasingly difficult to distinguish from human writing, motivating the development of automatic AI-generated text detectors. However, the robustness of such detectors under adversarial generation remains uncertain. This paper presents SHADE (Stochastic Human-like generation via Adversarial Detector Evasion), a reinforcement learning framework that formulates detector evasion as a policy optimization problem. Instead of applying post-hoc perturbations or prompting-based rewriting, SHADE fine-tunes an instruction-tuned LLaMA model with Group Relative Policy Optimization (GRPO), using feedback from a surrogate detector based on the PAN 2025 mdok system. Our experiments show that full fine-tuning with a small KL regularization penalty achieves $98.5\%$ surrogate evasion, compared to $1.5\%$ for the base model, while LoRA-based adaptation is substantially less effective under regularization. Linguistic analysis reveals that successful evasion is associated with shorter, simpler, and less lexically diverse outputs, suggesting that high detector evasion does not necessarily correspond to more human-like writing. In the official Voight-Kampff competition setting, our submissions ranked sixth and seventh, indicating that optimization against a single surrogate detector only partially transfers to unseen evaluation classifiers. These results highlight both the potential and limitations of reinforcement learning for adversarial AI-text generation and motivate more robust, multi-detector evaluation protocols for AI-generated text detection.

Evaluating an aligned language model by reading its answers assumes the answers carry the distinction the evaluator cares about. We introduce LLM endognostics, a white-box internal auditing framework designed to extract and causally manipulate latent knowledge within the residual stream. Applied to Minerva-7B-Instruct-v1.0 on 124 minimal prompt pairs over 12 categories of professional risk, behavioral evaluation fails on most of the set: the model acts identically on 63.7% of the pairs (95% CI [55.0%, 71.6%]), complying with or refusing both members. Yet, projecting the residual stream onto the vocabulary by a Jacobian lens reveals a statistically significant Contrastive Endognostic Margin, proving the model maintains robust risk differentiation internally. In a second protocol crossing 25 facts with five linguistic framings, we show that the model conforms to presupposed falsehoods in 72% of cases, despite representing the true entity in its latent layers. Surgically ablating the direction of the planted falsehood restores the correct answer in 11 of 25 suppressed cases (McNemar p = 0.0010), validated by blind human annotation (binary agreement kappa = 0.68). In contrast, an out-of-sample linear probe achieves 77% accuracy at layer 10, but its orthogonal ablation yields a 0% recovery rate. This establishes a fundamental theoretical dissociation: abstract linear representation does not imply causal control over verbalization. Our main contribution is the formalization of endognostics to prove that behavioral evaluation and internal reading systematically disagree in the common case, and that linear decodability is decoupled from causal control.

As LLM agents are exposed to hundreds to tens of thousands of skills, tools, and API functions, full-library prompting becomes costly, slow, and less reliable: each added candidate increases prompt tokens and latency, while longer candidate lists introduce more distractors for LLM selection. We present Toollery, a training-free candidate-compression framework for scalable LLM skill/tool selection. Following established document-side query expansion, Toollery generates user-intent queries from each skill/tool specification and builds a retrieval index that maps real user requests to compact candidate sets before final LLM decision-making. By treating high-level skills and atomic tools as selectable capabilities, Toollery can be applied to both skill libraries and tool registries. We evaluate Toollery on the roughly 79K-capability SkillRouter benchmark, BFCL-V4 with over 440 atomic tools, and 3,396 proprietary smart-cockpit requests over 220 tools. Across these settings, Toollery keeps online selection bounded to a compact top-K candidate set and improves recall over ordinary specification retrieval. At a fixed top-10 budget, Toollery improves end-to-end selection on the cockpit dataset, and maintains comparable AST Accuracy on BFCL-V4. These results support Toollery as a practical candidate-compression framework for large and evolving agent capability libraries, while showing that quality and cost gains depend on workload coverage and provider caching.

Quantization enables deployment of large language models on resource-constrained clinical edge devices, but its effect on clinical accuracy and safety remains understudied. We evaluate five 7-8B parameter models at FP16, GPTQ-INT8, and GPTQ-INT4 precision across five benchmarks: MedQA, MedMCQA, Med-HALT, a risk-stratified sample of HealthBench, and MedSafetyBench. The study jointly varies quantization bit width, model family, and clinical task type, with explicit risk stratification and safety measures. INT8 GPTQ is universally safe (max. degradation -1.9%-1.9%), while INT4 degradation is substantial and model-dependent: BioMistral-7B, clinically fine-tuned, loses 19.7% on MedMCQA, more than any general-purpose model, showing clinical fine-tuning does not confer compression robustness. MedMCQA degrades more than MedQA under INT4; Med-HALT is largely unaffected. On HealthBench's emergency-risk subgroup, Qwen2.5-7B degrades by 26.8% under INT4, suggesting high-risk scenarios are disproportionately vulnerable to compression. On MedSafetyBench, the model family dominates over precision (refusal rates range 10.2%-74.9% at FP16), though Qwen2.5-7B (-17.8%) and Meditron-7B (-28.3%) show substantial INT4 safety degradation; notably, Qwen2.5-7B is simultaneously the most accuracy-robust model, demonstrating that accuracy and safety robustness are independent properties. We additionally test two recovery methods, clinical calibration substitution and QLoRA fine-tuning, both producing the same trade-off: MedMCQA recovers while MedQA further degrades, indicating recovery strategies require task-specific validation rather than being assumed universally beneficial. These findings indicate INT8 is broadly safe for clinical deployment, while INT4 safety must be assessed per-model and per-task, and that safety alignment is determined primarily by instruction tuning rather than clinical domain adaptation.

Knowledge distillation can transmit unintended behavioral traits from a teacher model to a student through training data that appear semantically unrelated to those traits, a phenomenon known as subliminal learning. Although recent work has established this effect, its training dynamics and mitigation remain underexplored. We study subliminal learning in open-weight language models ranging from 1.5B to 8B parameters, covering the Qwen, Gemma, and Llama families in number-sequence and chain-of-thought settings. Rather than evaluating only final models, we track trait-related probabilities throughout fine-tuning and find that subliminal acquisition can be highly non-monotonic, with transient spikes, reversals, and trait-specific failures of transfer. We then introduce liminal training, an annealed KL-regularized fine-tuning method that constrains early drift from the base model. Across our experiments, liminal training substantially reduces subliminal trait acquisition while largely preserving task gains, outperforming paraphrasing and layer freezing as mitigation strategies. The effect also extends beyond animal preferences: in a French-language response-style experiment, liminal training suppresses language transfer while retaining much of the GSM8K improvement. Finally, we show that KL timing matters: early regularization is more effective than late regularization, and sweeping the regularization strength reveals an empirical trade-off between task learning and trait suppression.

Long multilingual conversational spoken question answering requires systems to balance long-range transcript semantics with sparse acoustic and speaker-sensitive cues. We present the Bairong system for the MLC-SLM 2026 Challenge, where a diarization-ASR front-end produces speaker-attributed transcripts and a dynamic evidence router constructs question-specific inputs for answer prediction. Instead of applying a fixed transcript-only or audio-only policy, the router infers the required evidence type and context scope from the question and answer options, and selects among full transcript context, local audio-text fusion, speaker-linked evidence, and compact global acoustic samples. This transcript-backbone design keeps discourse context available while activating audio only when it provides complementary evidence. Our Task 1 system achieves 25.70% and 18.44% tcpMER on the development and evaluation sets. For Task 2, the final system obtains 94.84% devel?opment accuracy, outperforming the full-transcript baseline by 1.68 points and the best audio-centric diagnostic system by 2.77 points. These results support dynamic question-aware routing as an effective evidence allocation strategy for conversational spoken QA.

Temporal Knowledge Graph Question Answering (TKGQA) requires answer inference from evidence that is both structurally valid and temporally admissible. Existing methods often leave anchor-event binding, temporal admissibility, and ordinal selection implicit in model reasoning, task-specific training, or similarity-driven retrieval, allowing locally relevant but invalid facts to enter the answer context. We formulate complex TKGQA as evidence-space control and propose SCoP (Structured Constraint Parsing), a constraint-centric framework that externalizes temporal decisions before answer inference. Instead of treating retrieved facts as admissible evidence by default, SCoP separates answer-seeking event patterns from temporal anchor events, conservatively grounds them to canonical TKG entities and relations, and translates temporal intent into executable constraints with optional ranking requirements. These constraints operate over normalized point and interval ranges, enabling deterministic filtering of structurally compatible candidates and producing a compact evidence space for generation. Experiments on MultiTQ and TimelineCronQ-R assess SCoP across timestamped point-fact and interval-oriented settings with richer temporal relations and ordering dependencies. Without task-specific parameter updates, SCoP achieves 0.825 Hits@1 on MultiTQ and 0.761 Hits@1 on TimelineCronQ-R, with gains on constraint-intensive question types. These results support explicit evidence-space control over unconstrained retrieval or implicit temporal reasoning.

Organizations in critical national infrastructure sectors must assess heterogeneous documents for sensitivity before routing or storage. Manual assessment is slow, inconsistent, and unscalable. Extending our prior leakage-controlled benchmark, BERT established the top single-encoder baseline (89.14% accuracy, 89.33% F1-score under 5-fold cross-validation on the Strategic 16K corpus). However, transformer baselines suffer from a structural limitation: fixed input length truncation discards evidence beyond the retained window-precisely where sensitive cables tend to be longest. We present Channel-Boosted MAS (CB-MAS) and instantiate it as IC-MAS (Iterative Consultation Multi-Agent System) to solve this without long-context computational costs. A Channel Critic Agent learns document-adaptive trust weights governing Gated Channel Boosting between two first-window encoders, while paired Consultation Agents iteratively exchange belief states to reconcile evidence from the beginning and end of long documents. IC-MAS holds computation constant regardless of document length by reconciling fixed windows in a compact representation space. Ablation studies show critic-controlled Channel Boosting provides the bulk of accuracy gains, while consultation recovers recall without precision collapse. Critic-Controlled Gated Channel Boosting with Max-Pool fusion and Blackboard Adaptive Consultation achieves 90.72% accuracy, 91.23% F1-score, 92.01% sensitive recall, and 90.46% sensitive precision, using about 54% less average computation than a fixed-round baseline. Gains over the single-encoder baseline are statistically significant (McNemar's test, p less than 0.000001; paired t-test). We include LIME/SHAP explainability, multi-agent evaluation, and an honest accounting of limitations.

Maxillofacial report generation from cone beam computed tomography is scored here by a composite objective placing 80% of its weight on a large language model judgement of factual entailment and 20% on lexical overlap, of which only the lexical fifth is visible during development. The grader's BLEU-4 and METEOR routines are reproduced in pure Python and match the reference to machine precision, and an offline entailment surrogate, which tells a report written for one patient from one written for another at an area under the curve of 0.987, makes the composite objective cheap enough to optimise directly. Over the 622-case public release, a report selected against the visible lexical ranking scores 0.2909, whereas one selected against the composite objective scores 0.4122, because pursuing n-gram overlap drives entailment precision from 0.522 down to 0.266. A 29 million parameter encoder fine-tuned on the release reaches a prevalence-weighted out-of-fold area under the curve of 0.486 over 985 statements, indistinguishable from the corpus prior, while nine numbers read from the image header reach 0.945 for mandible coverage and 0.872 for condyle coverage, and acquisition centre alone predicts sentence choice at 0.718 against 0.663 for the image-derived model, identifying dictation convention rather than anatomy as the quantity the lexical metrics reward. The delivered system emits eight unconditional statements and five gated on header geometry under polarity, laterality and tooth-level consistency constraints, and reaches METEOR 0.3542 over 50 held-out cases from an unseen centre. The dataset and code are available at https://github.com/GIND123/CBCT-Clinical-Reasoner

Orthodontic report generation from intraoral data is normally cast as multimodal captioning, yet the released Bite2Text scan pairs are supplied already registered in occlusion, which makes several core occlusal quantities directly measurable rather than inferable. The system reported here exploits that property: an anatomical frame is recovered per case from arch taper and arch closure instead of the stated RAS convention, which does not hold across the release, and each arch is reduced to an occlusal ridge profile in arch-angle coordinates yielding overbite, overjet, midline deviation, transverse overlap, crossbite extent, cusp interdigitation lag, and the occlusal curves in closed form. Gradient boosting maps 31 such measurements onto 13 template fields, a field being predicted only where patient-level cross-validation beats its own majority baseline, and a deterministic renderer emits the corpus six-part narrative; a ConvNeXt-Tiny classifier over the five standardised photographic views is fused per field, raising mean field accuracy from 0.601 to 0.683. Reimplementation of the challenge evaluator shows that its BLEU-4 and METEOR are local variants whose F-mean weights recall nine to one, that two clinicians agree on 47 percent of findings for the same patient, and that a constant report consequently outscores a genuine second clinician report by 0.165 captioning. Held-out scores reach BLEU-4 0.458 and METEOR 0.677 against intraoral scan references and 0.278 and 0.507 against photograph references, and the submitted system placed third in the ODIN 2026 Bite2Text test phase at 0.2680 and 0.4629, within 0.022 BLEU-4 of first, running on CPU in under ten seconds per case. The dataset and code are available at https://github.com/GIND123/ODIN_toothfairy4

No specification says how a governed autotelic AI agent organization, where agents pursue self-generated goals inside guardrails, should be designed and evaluated. We answer in two parts. First, we synthesize the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework from a document-based qualitative evidence synthesis of 321 sources, integrating agency, agile, platform, and governance theory into a runnable specification. Second, we probe a prompt-layer instantiation of GAMPO on CHI-Bench, a long-horizon healthcare benchmark, across open and frontier models. The result is a boundary condition: governance benefit is gated by a model's spare capacity and is domain- and model-specific. On capacity-constrained open models the full procedure yields no reliable benefit, whereas a single "verify your writes" sentence doubles task success (pass@1 2/20 to 4/20). At the frontier the same scaffold lifts prior-authorization 24% to 40% but nets zero on another model, a gap traced to a stable recommendation-override disposition. A second result refines the first: replacing the generic procedure with an answer-blind, per-task definition-of-done, keyed only to the case's own policy and published standards, never the hidden key, raises prior-authorization to 84% under best-of-five self-consistency (68% single-attempt, confirmed by a held-out board) and utilization-management to 44%, while care-management meets a subjective content-quality wall. The contribution is a named, auditable framework and capability-gated evidence that governance should be sized to spare capacity, and that at the frontier a case-grounded specification beats a uniform procedure. Findings are exploratory: partial instantiation, small per-cell samples (n = 5-25), and single trials.

This study examines the ability of large language models (LLMs) to predict the risk of weather-related forced outages in the distribution grid in a zero-shot framework, without labeled training data. The problem is formulated as a binary severity classification task across three forecast horizons (3h, 6h, 12h), using six years of outage records and high-resolution weather data for a utility service area in central Texas. Four zero-shot LLMs are benchmarked against two supervised classifiers across two input configurations: one using current weather observations and the other using weather forecast data. Results show that supervised models outperform LLMs on macro-F1 and precision, while newer LLM generations achieve competitive scores. Beyond accuracy, LLMs offer complementary strengths in actionable reasoning and geographic scalability, suggesting that combining them with supervised models may be the best practice.

Multi-agent LLM pipelines increasingly assign roles, including execution and verification, to models of different capability tiers. This is done because running a flagship model at every stage is expensive. Previous literature has established that verification stages are not always beneficial, but holds reviewer capability roughly fixed relative to the executor. We vary it. We replace the reviewer with models spanning a capability range down to one that cannot solve the problems at all, and measure the outcome of every individual rejection. This is done across a constant set of 100 olympiad mathematics problems. A cross-family mid-tier reviewer improves final accuracy by 12 percentage points, from 52 to 64 percent (p = 0.0005), with zero damaged answers. Same-model self-review attains the highest error-detection rate of any condition (0.85 recall) yet yields no significant gain: it rejects 2.1 times as often for a third the repair rate (15 against 43 percent, p = 0.0074) and falsely rejects 35 percent of its own correct answers against 2 percent for the cross-family reviewer (paired p = 0.000015). The low damage rate of self-review proves to be an artifact of revision inertia rather than reviewer quality: of 18 falsely rejected correct answers, the three where the executor complied all became wrong, while the fifteen it ignored survived unchanged. Below a capability floor the role becomes inert: our weakest reviewer changed zero of 100 final answers while doubling token cost. These findings describe a single executor-reviewer configuration on 100 problems and should be read as a controlled pilot rather than a general claim about verification stages.

Low-bit key–value (KV) caches reduce the memory required for autoregressive decoding, but the resulting quality loss depends on the model and quantizer. We keep the quantizer fixed and distill the floating-cache model's behavior into low-rank Q/K/V projection updates while the student executes a physically packed incremental cache. Across three seeds, 4-bit affine-cache adapters recover $54.24\%\pm2.47\%$ of the held-out perplexity gap on TinyLlama-1.1B and $75.96\%\pm4.04\%$ on Gemma-4-12B. On the same frozen NF4 Llama-3.1-8B base, one validation-selected run per quantizer recovers $60.42\%$ under KIVI K2V2 and $37.61\%$ under KVarN K4V2, while preserving 180-case associative retrieval. Gemma's score on an official 4K/8K RULER subset rises from 42.80 with the unadapted 4-bit cache to 48.33 after adaptation (46.15 floating), with substantial task heterogeneity. Finally, a 2-bit rank–token sweep reduces TinyLlama's 2-bit PPL from 576.10 to $11.4000\pm0.0059$ across three seeds, versus 10.3988 floating, but restores only 11–12 of 180 retrieval cases. These results show that low-rank projection adaptation can recover held-out quality across fixed cache formats, while perplexity recovery need not restore long-context retrieval.