Answer

Large language models are increasingly used as chemistry assistants, yet most chemistry benchmarks still score only final answers. This masks a critical failure mode: a model may output the correct molecule, product, or option while its reasoning violates chemical logic. Existing process-level evaluators are hard to scale because LLM judges and human step-level process annotation are costly, inconsistent, and vulnerable to hallucination. We introduce ChemCoTBench-V2, a rule-verifiable diagnostic benchmark for low-cost, auditable evaluation of structured, verifier-addressable chemical reasoning traces. It spans molecular understanding, molecule editing, molecular optimization, and reaction prediction, with 5,620 evaluation samples across 18 reporting tasks. Models must expose key intermediate steps in expert-designed templates, and those steps are checked with deterministic chemistry rules and, for closed-answer tasks, reference traces rather than another LLM judge. Open-ended molecular optimization is evaluated with oracle-verifiable state constraints rather than strict trace matching. The benchmark reports three separate signals: final-answer correctness, template adherence, and step-wise verifier correctness over expert-refined intermediate commitments. Experiments on frontier models reveal a persistent gap between final-answer success and structured-reasoning-state consistency: models often follow the requested format while failing chemical-step checks, or answer correctly with weak supporting reasoning. ChemCoTBench-V2 enables fine-grained model comparison and identifies the concrete step at which the trace first violates the verifier.

Biomedical question answering with large language models is commonly evaluated using answer accuracy, but answer accuracy alone does not indicate whether a model can produce parseable outputs, follow structured reliability instructions, recognize weak answer spaces, or avoid confident incorrect commitments. This paper presents HypothesisMed, an inference-time reliability pipeline for biomedical multiple-choice question answering. It combines direct, chain-of-thought, HypothesisMed-v3 prompting, and answer fusion. The final answer is selected by fusion, while HypothesisMed-v3 supplies SPACE labels and confidence information. SPACE labels mark the answer space as VALID, INCOMPLETE, or CONTRADICTED. We evaluate Qwen2.5-7B, Phi-4-mini, DeepSeek-R1-32B, and BioMistral-7B on MedQA, MedMCQA, and PubMedQA using 1,000 examples per dataset. The pipeline improves weighted accuracy over each model's best direct or chain-of-thought baseline while increasing parse and SPACE coverage. We also scale evaluation to Qwen2.5-7B and Phi-4-mini using 10,183 examples per model. Fusion improves Phi-4-mini accuracy from 0.4296 to 0.5192, while Qwen2.5-7B chain-of-thought remains slightly higher in answer accuracy. However, Qwen2.5-7B fusion achieves complete parse and SPACE coverage with much lower false commitment. A 12,000-example SPACE stress test shows answer-space diagnosis remains difficult, with SPACE accuracy of 0.3074 for Qwen2.5-7B and 0.4168 for Phi-4-mini. These results show that answer accuracy, parseability, structured reliability reporting, calibration behavior, and false-commitment behavior are separable capabilities. The main contribution is not a universal state-of-the-art claim, but a reproducible inference-time framework for evaluating biomedical question answering models as auditable workflow components under structured reliability constraints.

Multi-agent debate (MAD) is a promising strategy for improving LLM reasoning, but when agents converge on a shared answer, it is unclear whether that convergence reflects genuine deliberation or social compliance. We show that the conventional answer flip rate conflates three distinct mechanisms: spontaneous instability, stance-induced conformity, and reasoning-induced persuasion. Our three-source decomposition framework isolates each through controlled counterfactual conditions. In the primary MMLU-Pro setting, 37% of agent-question observations change under self-reflection alone, while robustness tests show substantial model-dependent instability across GPQA-Diamond and three model families; strict conformity is 29% in the primary setting and remains predominantly harmful across model replications (57-77% correct-to-wrong). A controlled information-gradient experiment reveals that even vacuous reasoning is associated with 20-39% error adoption among resistant agents, with reasoning-like presentation carrying substantial persuasive weight. Harmful conformity can be predicted from Round 0 features (AUC = 0.79), and risk-targeted intervention reduces it by 13.6 percentage points (p < 0.001). However, without correctness labels or self-reflection controls, reducing peer adoption does not improve accuracy, because harmful and beneficial influence cannot be distinguished.

Large language models (LLMs) are increasingly used for knowledge base question answering (KBQA), where answering requires selecting entities from a question-specific knowledge-graph subgraph. Yet LLMs are known to hallucinate across tasks, and KBQA is no exception: even when we provide a graph as the knowledge source, the model may rely on parametric knowledge instead of graph evidence or perform invalid reasoning over the given relations. Such hallucinated answer nodes can limit the practical deployment of KBQA systems, especially in high-stakes domains such as healthcare. We formulate hallucination detection in KBQA as an answer-node classification problem and propose a lightweight graph-based framework that treats the answering LLM as a black box. \methodname represents each KBQA instance as an augmented graph. It initializes node features with semantic representations of KG entities, marks topic entities and LLM-proposed answer nodes with learned vectors, and connect a virtual question node to the topic entities. A graph encoder then produces verification-oriented node representations, and a small MLP classifies each proposed answer node using its graph representation together with the question embedding. Experiments on WebQSP, ComplexWebQuestions, and PUGG show that our detector achieves the highest F1 on all three benchmarks ($82.0$, $87.4$, and $84.3$), outperforming LLM-as-judge and sampling-based baselines, while having $\sim305\times$ fewer parameters than the reference approaches. Beyond detection, the node-level feedback is actionable: when flagged answers are fed back to the KBQA system for iterative refinement, downstream KBQA F1 improves by $13.0$--$14.5$ points and Exact Match by $16.9$--$17.6$ points.

Self-play can train language models without external supervision. However, existing methods require rule-checkable answers, leaving open-ended tasks dependent on curated prompts or frontier-model judges. We introduce SCOPE, a data-free self-play framework for open-ended tasks that co-evolves two policies: a Challenger that generates document-grounded tasks, and a Solver that answers them through multi-turn retrieval. A frozen copy of the initial model serves as the self-judge, which writes task-specific rubrics from the source document and grades Solver responses against them. Across three 7-8B instruction-tuned models (Qwen2.5, Qwen3, OLMo-3), SCOPE improves open-ended performance by up to +10.4 points on eight benchmarks and matches or exceeds GRPO_data trained on ~9K curated prompts. Although trained only on open-ended tasks, SCOPE also improves held-out short-form QA by up to +13.8 points on seven held-out benchmarks, surpassing GRPO_data on all three models. Ablations show that co-evolving the Challenger is necessary to keep tasks near the Solver's frontier, that gains arise from improvements in both retrieval and synthesis with the relative contribution varying by task, and that rubric generation quality is the bottleneck for self-judging.

Retrieval-Augmented Generation (RAG) supplements a language model's input with retrieved documents, yet most RAG pipelines inherit retrieval components designed for human readers. How retrieved content should be represented when the consumer is a large language model (LLM) rather than a human is less well understood. Recent work has proposed transformations of retrieved content and identified properties that affect generation, but each examines a single transformation or property in isolation, leaving open which features of a document's representation matter most. We address this with a controlled comparison: holding retrieval fixed, we vary only the representation of retrieved documents, comparing an original baseline against thirteen transformations spanning selection, summarisation, and reformulation, in query-dependent and query-independent variants. Across these fourteen representations we measure question-answering accuracy for four generators, and for each representation we also measure answer retention: whether a known answer-bearing document still supports its answer after transformation. We find that answer retention is the primary determinant of generator accuracy; notably, when retention is high, a representation's wording, structure, length, and query-dependence have limited effect. This suggests that accuracy gains attributed to specific mechanisms in prior work may be partly explained by how well those mechanisms preserve answer-bearing content, an attribution that cannot be settled without controlling for retention.

Spatial reasoning is a fundamental capability for vision-language models (VLMs) deployed in real-world environments. However, visual observations are inherently limited representations of a 3D world: occlusion can render objects invisible, and perspective can make geometric properties misleading. Despite this, existing spatial reasoning benchmarks typically assume that observations are sufficient and reliable, focusing on whether models produce correct answers rather than whether they recognize when a question cannot be answered and what additional observations would be needed. In this work, we challenge this assumption by constructing a controlled evaluation framework, SpatialUncertain, and introducing two types of observation challenges: (1) occlusion, which hides target information, and (2) perspective ambiguity, which produces misleading visual cues. For each configuration, we design spatial questions that are answerable under clean observations but require abstention under the introduced challenges. We further evaluate whether models can identify which additional viewpoints would resolve perspective ambiguity. Our results across a diverse set of frontier open- and closed-source VLMs reveal two consistent failure modes. First, models are prone to overconfident answering, attempting to solve spatial reasoning tasks even when visual evidence is incomplete or misleading, with average accuracy around 30% under occlusion and below 10% under perspective ambiguity. Second, even when additional views are available, some models perform near random chance in identifying which would provide reliable evidence. Together, our findings call for moving beyond answer correctness toward evaluating whether models know when to abstain and how to seek reliable evidence.

Many open-ended instructions have multiple valid answers that users can benefit from seeing, but post-training often narrows an LLM's output space toward a small set of canonical responses. We introduce REDIPO, an offline DPO data-construction pipeline for recovering distinct valid answer modes while preserving the alignment benefits of the instruct model. For each prompt, REDIPO samples responses from both base and instruct models, rewrites base-model responses with the instruct model, filters candidates for safety and instruction-following quality, and builds preference pairs that favor marginally diverse responses among candidates with similar instruction-following reward. Across Qwen3-4B, OLMo-3-7B, and LLaMA-3.1-8B, REDIPO improves NoveltyBench distinct_k by 134%, 33%, and 44% relative to the instruct checkpoints, while DivPO changes diversity by 0%, -6%, and -4% on the same models. These gains largely maintain MTBench, IFEval, and Arena-Hard performance, and reduce direct-category HarmBench attack success rate. Ablations show that marginal-diversity pair selection and base-response rewriting drive the diversity gains, while filtering and quality-bounded pairing help maintain alignment. Overall, our results show that diverse valid answers from base-model generations can be reintroduced through carefully constructed preference data while retaining the alignment benefits of post-training. We release our code and data at https://github.com/vsamuel2003/ReDiPO.

Decoding visual content from fMRI signals recorded while a person views images, and specifically answering questions about the seen images, is a long-standing challenge. While significant progress has been made in recent years in visual question answering (VQA) from fMRI, performance remains limited. Moreover, although recent models can make increasingly accurate predictions, they have rarely been used as tools for understanding the structure of visual representations in the brain. We present Brain-IT-VQA, a framework for visual question answering from fMRI. Building on the Brain Interaction Transformer (Brain-IT), our method decodes language tokens from brain activity and integrates them with a language model to answer visual questions. Our model substantially outperforms previous fMRI-based captioning and VQA approaches. We further introduce NSD-VQA, a new dataset and benchmark for visual question answering from fMRI. Unlike existing image-fMRI VQA datasets, which typically provide only a few broad and weakly controlled questions per image, NSD-VQA provides on average 20 question-answer pairs per image across 20 controlled question categories that disentangle multiple levels of visual understanding. This enables more reliable and interpretable evaluation despite limited fMRI test data. Together, Brain-IT-VQA and NSD-VQA provide both a strong predictive framework and a tool for studying brain representations. Using this benchmark, we quantify which forms of visual and semantic information can be reliably decoded from fMRI responses to natural images. We further analyze the contributions of different brain regions across question types.

Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers. Under sustained adversarial pressure we find a previously undocumented failure mode: the chain-of-thought stays factually correct from first turn to last while the emitted answer flips wrong. We call this unfaithful capitulation (UC) and isolate it with a $2\times 2$ latent-versus-behavioral framework that flip-rate metrics and single-turn faithfulness probes both miss. Across three datasets (MT-Consistency, MMLU-Pro, GSM8K), the latent-correct rate at the behavioral flip clusters near 50% in think mode and collapses to 11-15% under no_think -- paired, within-model causal evidence that reasoning creates the gap. Across models the effect tracks the reasoning channel (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it). An independent GPT-4o judge corroborates $86\%$ of UC labels; a token-level probe shows the answer-slot argmax is correct in $84\%$ of UC cells; and a naive trace-anchored defense backfires. We release all trajectories, traces, and judge labels.

We highlight a failure mode of large reasoning models on questions with insufficient information: models may recognize that a problem is under-specified, yet still continue reasoning and produce unsupported final answers instead of abstaining. We formalize this mismatch as the detection-to-abstention gap, where detected insufficiency fails to translate into final abstention. This gap is especially concerning in high-risk domains such as medical AI, where answers based on incomplete evidence can be more harmful than refusal. To close this gap, we propose Judge-Then-Solve (JTS), a trajectory-level reasoning-control framework that trains models to make an explicit answerability commitment before solution generation. Rather than treating abstention as a final-answer style, JTS casts it as a control decision: the model either proceeds to solve or terminates early based on its answerability judgment. We instantiate this policy through supervised warm-up and missing-premise reinforcement learning with consistency and length-shaping rewards. Experiments on dense and MoE reasoning models show that JTS substantially improves reliable abstention across datasets and pushes Abstention@Detection (A@D) to near-saturation, indicating that models not only detect missing information but also act on that detection. By terminating unanswerable trajectories immediately after the answerability judgment, JTS reduces unnecessary reasoning and improves inference efficiency when continued deliberation would amplify unsupported assumptions. We also observe that missing-premise training can alter reasoning behavior on difficult but answerable problems, reducing unproductive self-reflection. These results suggest that abstention under insufficient information is a key form of reasoning control for deploying reasoning models safely and efficiently.

CASTLE 2026 asks 185 multiple-choice questions over 600+ hours of synchronized multi-view egocentric video. We explore two approaches on top of a shared multimodal preprocessing layer, including per-person timelines, speaker-resolved transcripts, and multi-VLM caption ensembles. Approach A, SVA: Search-Verify-Answer, is a three-stage pipeline that hierarchically narrows to a primary window, verifies sub-windows with a VLM under four anti-confabulation rules, and fuses evidence with an LLM judge under an evidence-priority hierarchy. Approach B, TMKG: Temporal-Multimodal-Knowledge-Graph, is the contrast: it builds a temporal multimodal knowledge graph, locates a primary cell via graph search, and produces the final answer with a single grounded VLM. SVA reaches a leaderboard accuracy of 0.50 and is our final challenge submission; TMKG reaches 0.35.

Modern retrieval-augmented generation(RAG) deployments increasingly rely on caching to reduce token cost and time-to-first-token(TTFT). Prefix-level KV reuse is now standard in serving stacks such as vLLM, and chunk-level and position-independent reuse have been pushed further by recent systems(RAGCache, TurboRAG, CacheBlend, EPIC, ContextPilot, PCR, LMCache). Output-level semantic answer caches, by contrast, remain fragile: similar prompts can map to different correct answers, retrieved evidence drifts as the corpus is updated, and adversarial collision attacks have been shown to hijack cached responses. We argue that the right framing for cached answer reuse is not how to reuse faster but when reuse is safe. We propose GroundedCache, an evidence-validated cache router that admits a cached answer only when 4 cheap gates simultaneously hold: query similarity, retrieved-evidence overlap, source-version validity, and lexical (or judge-based) support of the cached answer by the freshly retrieved evidence. We build a six-regime workload that stress-tests cache safety rather than only hit rate, and introduce an operator-facing metric, the unsafe-served rate (USR), fraction of all queries that received a wrong cached answer. Across 2 datasets and 12,000 real-LLM generations(Qwen2.5-7B-Instruct on vLLM with Automatic Prefix Caching), GroundedCache drives USR to 0.0% on every HotpotQA regime(vs. 15-35% under naive caching) and to 1.5% on mtRAG document drift(vs. 51.5%), a 34x reduction on the design-point adversarial regime and 3-10x reductions across the other mtRAG regimes, while end-to-end p50 latency stays within 1.04-1.07x of a no-cache RAG baseline. A per-gate ablation isolates the lexical support gate as the load-bearing safety mechanism on both datasets, with the remaining gates providing defense-in-depth at near-zero cost. We release the implementation, workload, and evaluation harness.

Retrieval-Augmented Generation (RAG) systems for question answering typically retrieve evidence by semantic similarity between the query and document chunks. While effective for unstructured text, this approach is less reliable on semi-structured corpora where answering may require exact filtering, aggregation, or exhaustive retrieval over structured attributes across multiple documents. Symbolic approaches support such operations, but they are often brittle on noisy natural-language corpora. We address this gap with DualGraph, a RAG framework that represents documents through two complementary views: a Textual Knowledge Graph for semantic retrieval and a Symbolic Knowledge Graph for symbolic querying over typed subject--predicate--object triples. Building on these two components, we provide multiple strategies for selecting or combining semantic and symbolic evidence.We also introduce SpecsQA, a benchmark from a commercial shopping website with semi-structured product documents and manually curated questions spanning open-ended and specification-oriented retrieval. Experiments show that DualGraph consistently outperforms state-of-the-art dense-retrieval, GraphRAG, symbolic, and table-oriented baselines across question types.Code and data are available at https://github.com/corneliocristina/DualGraphRAG.

LLM training increasingly relies on teacher-generated supervision, from synthetic responses to reasoning traces and tool-use demonstrations. Current practice often chooses the highest-performing teacher to generate student training data, implicitly treating teacher test performance as a proxy for teaching quality. We show that this assumption can fail: even when multiple teachers provide correct answers to the same question, the answer from the strongest teacher is not necessarily the best supervision for a given student. To address this gap, we propose Student-Centric Answer Sampling (SCAS), a framework that selects from verified teacher-generated answers according to their estimated student-centric learning cost. Motivated by a token-wise gradient decomposition, we derive an efficient forward-only proxy for this cost and use it to guide answer selection during training. Experiments across 30 teacher models, 6 student base models, and 6 tasks show that SCAS consistently improves student performance, suggesting that effective distillation should prioritize supervision matched to the current student rather than teacher strength alone.

Evidence construction--the stage that determines which passages reach the language model before generation begins--is evaluated paradigm by paradigm, leaving practitioners with no principled way to diagnose which organization strategy fails, where, or why. We introduce AuthTrace, a diagnostic benchmark built on thematically dense single-author corpora where near-miss distractors share style, topic, and vocabulary with the required evidence. AuthTrace provides explicit quoted evidence, exact fan-in annotation, and a unified pack-level protocol measuring evidence recall, evidence precision, and answer correctness. A fan-in gradient--the number of source documents required to support the answer--serves as the primary diagnostic axis, enabling controlled comparison across retrieval, memory, graph, and structured-evidence paradigms. Evaluating eight systems across two QA models, we find that evidence recall is the strongest observed predictor of answer correctness under the primary reader-judge pair (r = 0.96); most failures stem from missing evidence rather than answer synthesis. Fan-in further exposes paradigm-specific collapse patterns: flat retrieval degrades 2-3x faster than thematically organized evidence construction. These results show fan-in decomposition to be a reusable diagnostic lens for identifying where evidence-construction systems fail and which paradigm best serves a given workload.

Before an LLM agent can use a tool, a retrieval system must decide which candidate tools to show to the agent. How long should that shortlist be? Show too many tools and the model struggles to choose. Show too few and the correct tool may not appear. Most systems apply a fixed shortlist size to every query, but no standard metric exists to evaluate whether that size was appropriate. We treat the number of tools shown to an LLM agent as the object of evaluation and we apply Bits-over-Random (BoR), a chance-corrected metric that asks whether success at a given depth is better than what random selection would achieve at that same depth. We evaluate BoR across three tool-selection benchmarks, multiple scorers, and registries ranging from 20 to 3,251 tools. We then turn the same principle into a reinforcement learning (RL) reward for choosing tool shortlist depth per query. The RL agent is deliberately simple, serving as a probe of the metric rather than a proposed system. As the shortlist grows, random chance of including the correct tool rises, so the reward naturally decreases, reducing the need for an engineered depth penalty. On BFCL (370 tools), the learned policy nearly matches the coverage of showing 50 tools ($90.3\%$ vs $90.8\%$) while presenting only 7 on average. On ToolBench (3,251 tools), a fixed shortlist of 5 tools achieves higher aggregate coverage ($64.7\%$ vs $61.9\%$) but finds nothing on hard queries (correct tool ranked 6th-20th). The BoR agent finds $16.7\%$ on those same queries by searching deeper. Downstream validation with Claude Sonnet 4.6 indicates that shorter adaptive lists also improve the LLM's ability to select the right tool: $93.1\%$ versus $87.1\%$ when always shown 5 tools, widening to $76.8\%$ vs $60.9\%$ on medium-difficulty queries where the correct tool is present but not ranked first.

Large Language Models (LLMs) are increasingly deployed as autonomous agents that reason, use tools, and act over multiple steps. Yet most hallucination benchmarks still evaluate only the final output, missing failures that originate in intermediate Thought-Action-Observation steps. We present Trajel, a dataset and evaluation framework for auditing trajectory-level hallucinations in multi-agent industrial workflows. Trajel introduces a five-type hallucination taxonomy (factual, referential, logical, procedural, and scope-based) over expert-annotated agent traces from AssetOpsBench. We benchmark supervised detection models at the subtask, trajectory, and long-context levels. Our results show that the most common failure modes are missed by existing benchmarks, that nearly half of hallucinated trajectories involve multiple types at once, and that automated detectors with high binary accuracy still misclassify the subtlest types. Trajectory-aware detection significantly outperforms standard post-hoc verification, making taxonomy-grounded evaluation necessary for safer agentic deployment.

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

Hallucination is often viewed as a direct consequence of missing knowledge: a model answers incorrectly when the correct answer is absent from its generation-time distribution, and correctly when it is present. We test this assumption by introducing a semantic notion of answer availability that aggregates token-level variants expressing the same answer concept, and asks whether the correct concept is already available at the moment the model commits to an answer. Across Qwen and Llama models from 0.8B to 72B in both Instruct and Base variants, 16-47% of Instruct hallucinations occur with substantial probability mass already on the correct concept, and the rate rises monotonically with scale. Comparing such failures against correct generations with matched semantic support, the distinguishing factor is not whether the correct concept is represented, but how its probability is distributed: correct generations concentrate mass on a single surface form, hallucinations disperse it across alternatives. The same sharpening asymmetry extends across multi-token generation and is detectable in pre-generation hidden states. Together, these results identify a single mechanism: instruction tuning sharpens answer commitment with scale, making helpfulness and confident hallucination two consequences of the same underlying disposition.

Pairwise image quality assessment (IQA) in professional photography requires a model not only to identify the preferred image between two candidates, but also to provide convincing and image-grounded reasoning. In the NTIRE 2026 RAIM challenge, this requirement is further emphasized by jointly evaluating preference prediction and rationale generation. To address this task, we propose iDiff, an Interpretable Difference-aware framework for pairwise image quality assessment. Our method adopts a dual-branch design consisting of an Answer Model and a Thinking Model. The Answer Model performs robust preference prediction by explicitly decomposing each sample into left/right global and local views, followed by content-aware specialization for person and scene images and ensemble-based aggregation across backbones. The Thinking Model focuses on rationale generation and is progressively enhanced with expert-style templates, multi-source quality features, and answer-aware supervision conditioned on the Answer Model prediction. In this way, iDiff jointly models discriminative decision making and structured explanation, improving both robustness and interpretability. Extensive experiments demonstrate the effectiveness of the proposed framework on both accuracy and reasoning-quality metrics. Our method achieved first place in the NTIRE 2026 RAIM challenge, showing the effectiveness of integrating explicit difference modeling with structured multimodal reasoning for pairwise IQA.

Small vision-language models (2-8B) are well-suited for clinical deployment due to privacy constraints, limited connectivity, and low-latency requirements favouring on-device or on-premise inference. However, their limited capacity exacerbates the generation of plausible but incorrect outputs. We extend game-theoretic decoding, previously restricted to text-only, closed-ended NLP tasks, to vision-language models for open-ended Medical VQA. We introduce a semantically aware Wasserstein stopping criterion that replaces lexical order matching, enabling convergence based on semantic consensus among near-synonymous candidate answers and avoiding unnecessary iterations caused by clinically equivalent ranking swaps. On VQA-RAD and PathVQA, we obtain consistent, statistically significant improvements over greedy and discriminative baselines. On VQA-RAD, we improve Qwen3-VL-2B by +3.5 percentage points (p < 0.01), surpassing the greedy 4B model, with similar trends at larger scales. On PathVQA, Gemma-3-4B with BDG matches MedGemma-4B under greedy decoding despite no domain-specific fine-tuning. At accuracy parity with classic BDG, the Wasserstein criterion reduces average convergence iterations by approximately 20%, improving inference efficiency while preserving the game-theoretic equilibrium behaviour. Code is available at https://github.com/luca-hagen/ Wasserstein-BDG-medical-VQA.

Large Reasoning Models (LRMs) achieve strong performance by generating long chains of thought (CoT), but often overthink, continuing to reason after a solution has already stabilized and thereby wasting tokens and increasing latency. Existing inference-time early-exit methods rely primarily on answer-level signals, such as confidence or trial-answer consistency, to decide when to stop. However, these signals mainly reflect answer readiness rather than reasoning convergence: they may trigger before the model has finished exploring or self-correcting, causing premature exits that can degrade final-answer accuracy and leave the retained reasoning chain semantically incomplete. We identify reasoning-level semantic redundancy as a complementary signal for semantic-preserving early exit: when successive steps no longer add novel progress and instead revisit established conclusions, the reasoning trajectory has likely converged. Building on this insight, we propose PUMA, a plug-and-play framework that combines a lightweight Redundancy Detector with answer-level verification. The detector flags semantically redundant candidate exits, while verification confirms whether stopping is safe, allowing PUMA to remove redundant continuation while preserving both answer accuracy and a coherent reasoning prefix. Across five LRMs and five challenging reasoning benchmarks, PUMA achieves 26.2% average token reduction while preserving accuracy and retained CoT quality. Additional experiments on code generation, zero-shot vision-language reasoning, and learned stopping-policy internalization further demonstrate that reasoning-level redundancy is a robust, transferable, and learnable signal for efficient reasoning. Our code is available at \url{https://github.com/giovanni-vaccarino/PUMA}.

Evaluations of unlearning on reasoning models sometimes show a bypass pattern. The answer side looks unlearned, but the model's own thinking trace keeps emitting the forgotten content, and the gap is taken as evidence that the weights still remember. We audit this reading on DeepSeek-R1-Distill-Qwen-7B with LoRA-memorized fictional authors and NPO unlearning, conditioned on a six-token canary head. On one seed, swapping the thinking trace for a short non-canary prefill on the same weights drops the answer rate by as much as the bypass gap itself, whether the prefill mimics the training template or not. On a second seed the bypass gap shrinks rather than vanishing, and the prefill swap reverses direction and brings the answer rate to ceiling. A positive parser-split bypass gap thus does not by itself identify hidden weight-level memorization, and does not rule it out either. On a different distillate the same metric flips sign because the parser cannot find the closing tag. We recommend a decode-time template swap as a cheap sanity check alongside the canonical audit.

Hallucination detection in large language models (LLMs) requires balancing accu racy, efficiency, and robustness to distribution shift. Black-box consistency methods are effective but demand repeated inference; single-pass white-box probes are effi cient yet treat answer representations in isolation, often degrading sharply under domain shift. We propose QAOD (Question-Answer Orthogonal Decomposition), a single-pass framework that projects away the question-aligned direction from the answer representation to obtain a question-orthogonal component that suppresses domain-conditioned variation. To identify informative signals, QAOD further selects layers via diversity-penalized Fisher scoring and discriminative neurons via Fisher importance. To address both in-domain detection and cross-domain generalization, we design two complementary probing strategies: pairing the or thogonal component with question context yields a joint probe that maximizes in-domain discriminability, while using the orthogonal component alone preserves domain-agnostic factuality signals for robust transfer. QAOD's joint probe achieves the best in-domain AUROC across all evaluated model-dataset pairs, while the orthogonal-only probe delivers the strongest OOD transfer, surpassing the best white-box baseline by up to 21% on BioASQ at under 25% of generation cost.

Long video question answering requires locating sparse, time-scattered visual evidence within highly redundant content. Although current MLLMs perform well on short videos, long videos introduce long-horizon search and verification, which often necessitates multi-turn, agentic interaction. We show that existing LVU agents can exhibit "evidence misalignment": they produce correct answers that are not supported by the retrieved or inspected evidence. To characterize this failure, we introduce two diagnostics (temporal groundedness and semantic groundedness) and use them to reveal two pressures that amplify misalignment: prompt pressure from shared-context saturation at inference time and reward pressure from outcome-only optimization during training. These findings point to a structural root cause: the coupled agent paradigm conflates long-horizon planning with answer authority. We therefore propose the decoupled planner-inspector framework, which separates planning from answer authority and gates final answering on pixel-level verification. Across four long-video benchmarks, our framework improves both answer accuracy and evidence alignment, achieving 55.1% on LVBench and 62.0% on LongVideoBench while producing interpretable search trajectories. Moreover, the decoupled architecture scales consistently with increased search budgets and supports plug-and-play upgrades of the MLLM backbone without retraining the planner. Code and models are available at https://github.com/Echochef/VideoSEAL.

Solving complex long-horizon tasks requires strong planning and reasoning capabilities. Although datasets with explicit chain-of-thought (CoT) rationales can substantially benefit learning, they are costly to obtain. To address this challenge, we propose Hindsight Hint Distillation (HHD), which only requires easy-to-obtain question-answer pairs without CoT annotations. Inspired by how human teachers use student mistakes to provide targeted guidance, HHD synthesizes hindsight hints from the model's own failed self-rollouts and uses them to scaffold on-policy rollouts that successfully complete the tasks. The model then self-distills these scaffolded trajectories and generalizes to new problems without hint guidance. Experiments show that HHD significantly outperforms iterative RFT and trajectory-synthesis baselines, achieving an absolute improvement of 8% on SWE-bench Verified, while all baselines improve by only around 2%. Notably, the reasoning strategies induced by HHD generalize effectively to out-of-distribution tasks, yielding the largest gains on SWE-bench Multilingual despite no training on multilingual data. These results demonstrate that HHD can effectively synthesize expert-like reasoning from CoT-free data and substantially improve long-horizon performance.

Multimodal large language models (MLLMs) have heterogeneous strengths across OCR, chart understanding, spatial reasoning, visual question answering, cost, and latency. Effective MLLM routing therefore requires more than estimating query difficulty: a router must match the multimodal requirements of the current image-question input with the capabilities of each candidate model. We propose LatentRouter, a router that formulates MLLM routing as counterfactual multimodal utility prediction. Given an image-question query, LatentRouter extracts learned multimodal routing capsules, represents each candidate MLLM with a model capability token, and performs latent communication between these states to estimate how each model would perform if selected. A distributional outcome head predicts model-specific counterfactual quality, while a bounded capsule correction refines close decisions without allowing residual signals to dominate the prediction. The resulting utility-based policy supports performance-oriented and performance-cost routing, and handles changing candidate pools through shared per-model scoring with availability masking. Experiments on MMR-Bench and VL-RouterBench show that LatentRouter outperforms fixed-model, feature-level, and learned-router baselines. Additional analyses show that the gains are strongest on multimodal task groups where model choice depends on visual, layout-sensitive, or reasoning-oriented requirements, and that latent communication is the main contributor to the improvement. The code is available at: https://github.com/LabRAI/LatentRouter.

Recently there has been an increasing interest in incorporating intensional'' functions in answer set programming. Intensional functions are those whose values can be described by other functions and predicates, rather than being pre-defined as in the standard answer set programming. We demonstrate that the functional stable model semantics plays an important role in the framework of Answer Set Programming Modulo Theories (ASPMT)'' -- a tight integration of answer set programming and satisfiability modulo theories, under which existing integration approaches can be viewed as special cases where the role of functions is limited. We show that ``tight'' ASPMT programs can be translated into SMT instances, which is similar to the known relationship between ASP and SAT.

Large Language Models (LLMs) voice assistants are commonly built as cascaded Automatic Speech recognition (ASR) to LLM systems, where recognition errors can distort user intent. Dislikes may also arise from ambiguous, out-of-domain, or non-request turns, making it hard to isolate ASR effects. We release WASIL (it denotes connection or linking in Arabic): in-the-wild Arabic spoken interaction prompts with audio, ASR hypotheses, assistant responses, and explicit like/dislike feedback (8,529 turns; 14.2% dislikes), plus a 2,000-turn test set covering Modern Standard Arabic (MSA) and four major dialects with their labels. We provide low-cost gold transcripts via multi-ASR agreement-guided post-editing and annotate answerability (answerable, ambiguous/needs-clarification, unsupported, not-a-request/noise) to separate intrinsic unanswerability from ASR-induced degradation. Finally, we describe scalable reference-free evaluation of responses from ASR vs. gold transcripts using multi-judge LLM scoring.

A standard technique for scaling inference-time reasoning is Self-Consistency, whereby multiple candidate answers are sampled from an LLM and the most common answer is selected. More recently, it has been shown that weighted majority voting (e.g. Confidence-Informed Self Consistency (CISC)), which assigns a confidence value to each candidate answer and chooses the answer with the largest accumulated score, tends to be more accurate on a wide range of popular benchmarks. In practice, weighted majority voting necessitates calling a critic LLM on each candidate's reasoning trace to produce the answer's confidence score. This secondary series of LLM calls greatly increases the overhead and cost of weighted majority voting, despite its potential performance benefits. To reduce this expense, we propose VecCISC, a lightweight, adaptive framework that uses a measure of semantic similarity to filter reasoning traces that are semantically equivalent to others, degenerate, or hallucinated, thus decreasing the number of candidate answers that must be evaluated by the critic. To ensure adequate experimental thoroughness, we evaluate VecCISC on five challenging, widely-adopted datasets spanning the domains of mathematics, chemistry, biology, commonsense reasoning, and the humanities. Our results demonstrate that VecCISC reduces the total token usage by 47%, while maintaining or exceeding the accuracy of CISC.

Modern reasoning language models generate dense, sequential chain-of-thought traces implicitly assuming that every token contributes and that steps must be consumed in order. We challenge both assumptions through a systematic intervention pipeline--removal, masking, shuffling, and noise injection--applied to model-generated reasoning chains across three models and three benchmarks. Our findings are counterintuitive on three dimensions. Order: Does the sequential order of a reasoning chain matter for answer extraction? No--line-level shuffling reduces accuracy by less than 0.5 pp; word-level shuffling retains 62%-89% accuracy; only token-level shuffling collapses to near zero. Pretrained-only and instruction-tuned variants exhibit near-identical tolerance (78.67% vs. 78.00% under line shuffling), indicating order-independence originates from pretraining rather than reasoning-specific fine-tuning. Dense: Is all the information in a reasoning chain important for answer extraction? No--masking numeric digits collapses accuracy to exactly 0%, while masking alphabetic prose improves accuracy by 4.7 pp. Robustness: Is a reasoning chain that is both order-shuffling and non-dense still robust? Yes--the most aggressively reduced representation (all natural language removed, lines arbitrarily shuffled) still achieves 83% accuracy, and injecting false answers at 3x true-answer frequency leaves accuracy unchanged (83.3%->83.3%), falsifying a frequency-based extraction account. These results establish that answer extraction operates on a sparse, order-insensitive, and structurally robust informational substrate, opening paths toward parallelized and token-efficient reasoning generation.

Artificial intelligence offers powerful new tools for scientific discovery, but the interaction paradigms required to effectively harness these systems remain underexplored. In this paper, we present findings from a formative user study with 11 expert mathematicians who used AlphaEvolve, an evolutionary coding agent, to tackle advanced problems in their fields of expertise. We identify and characterize a distinct workflow we term intentmaking, the iterative process of discovering, defining, and refining one's experimental goals through active system interaction. We frame this as a natural extension to sensemaking, the cognitive process of building an understanding of complex or novel data. We suggest that users enter a cycle of intentmaking (defining and updating their experiment) and sensemaking (interpreting the results) which repeats many times during the course of an investigation. Our documentation of these themes suggests an approach to designing AI tools for scientific discovery that goes beyond the existing question/answer model of many current systems, treating them as collaborative instruments rather than opaque black-box assistants.

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Auditing language-model outputs often requires more than judging correctness: an auditor may need to identify which source document most likely supports the knowledge expressed in a response. We study this as pinpoint provenance: given a prompt, a target-model response, and a candidate corpus, rank the documents that best support the response. We introduce FakeWiki, a controlled benchmark of 3,537 fabricated Wikipedia-style articles designed to preserve ground-truth provenance while weakening lexical shortcuts. FakeWiki includes QA probes, source-preserving paraphrases, retro-generated variants, hard anti-documents that remain topically similar while removing answer-critical facts, and five query conditions: clean prompting plus four jailbreak-inspired transformations. We evaluate seven retrieval baselines, a training-free activation-steering retrieval-fusion method, SteerFuse, and a supervised contrastive provenance ranker, ScoringModel. ScoringModel maps response and document features into a shared space and is trained with InfoNCE using in-batch, retrieval-mined, and anti-document negatives. Across nine open-weight instruction-tuned LLMs and five query conditions, ScoringModel improves mean Recall@10 from 35.0 for the strongest retrieval baseline to 52.2, without inference-time fusion, and wins 41/45 model-by-condition cells. SteerFuse is usually second-best despite requiring no supervised training, showing that activation-space evidence can efficiently complement text retrieval. On jailbreak-inspired transformed queries, ScoringModel improves Recall@10 by 15.7 points on average over the best baseline. Overall, our work shows that robust training data attribution requires evaluation settings that separate true answer support from topical or lexical resemblance.

Self-consistency detects hallucinations by generating multiple sampled answers to a question and measuring agreement, but this requires repeated decoding and can be sensitive to lexical variation. Semantic self-consistency improves this by clustering sampled answers by meaning using natural language inference, but it adds both sampling cost and external inference overhead. We show that first-token confidence, phi_first, computed from the normalized entropy of the top-K logits at the first content-bearing answer token of a single greedy decode, matches or modestly exceeds semantic self-consistency on closed-book short-answer factual question answering. Across three 7-8B instruction-tuned models and two benchmarks, phi_first achieves a mean AUROC of 0.820, compared with 0.793 for semantic agreement and 0.791 for standard surface-form self-consistency. A subsumption test shows that phi_first is moderately to strongly correlated with semantic agreement, and combining the two signals yields only a small AUROC improvement over phi_first alone. These results suggest that much of the uncertainty information captured by multi-sample agreement is already available in the model's initial token distribution. We argue that phi_first should be reported as a default low-cost baseline before invoking sampling-based uncertainty estimation.

Large language models often fail at simple counting tasks, even when items to count are in the prompt. We investigate whether this failure occurs because transformers do not represent counts internally, or because they cannot convert representations to the correct output tokens. Across three model families: Pythia, Qwen3, and Mistral, ranging from 0.4B to 14B parameters, we find evidence for the second explanation. Linear probes recover the correct count from intermediate layers with $R^2>0.99$, showing that the information is present. However, the internal directions that encode counts are nearly orthogonal to digit-token output-head rows ($|\cos| \leq 0.032$). In other words, the model stores the count in a form that the digit logits do not naturally read out. We localize this failure with two interventions. Updating only the digit rows of the output head (36,864 parameters) substantially improves constrained digit prediction (60.7--100.0% on four tasks), but it does not fix unconstrained generation (0%); we do not claim that digit-row repair fixes open-ended text. By contrast, small LoRA on attention Q/V (7.67M parameters) improves upstream routing and achieves 83.1%$\pm$7.2% in true greedy autoregressive generation (deployable fix). Logit-lens at layer 35 (entity counting; correct-digit rank): (i) median over 3 seeds drops from order-$10^4$ to 1; (ii) seed 42 shows $54{,}332 \to 838$ (median top-1 while one seed stays far below). Norm, logit-lens, and cross-task analyses generalize the bottleneck to counting, addition, and list length; nulls on MMLU and GSM8K and limited DROP transfer. These results identify counting failure as a geometric readout bottleneck, not an internal-representation failure: the model knows the count but the output pathway is misaligned with tokens needed to express it.

A reliable language model should be able to signal, prior to generation, when a query falls outside its knowledge. We investigate whether representation geometry can provide such a pre-generation signal by measuring the deviation of hidden states from an answerable reference set, requiring no labeled failure data and no access to model outputs. Across three instruction-tuned models (Llama 3.1-8B, Qwen 2.5-7B, and Mistral-7B-Instruct) and three prompt forms (Math, Fact, Code), we find that geometry primarily encodes task form. Within mathematical prompts, unanswerable inputs consistently deviate from the answerable centroid, yielding strong separation (ROC-AUC 0.78-0.84). This single-pass pre-generation signal outperforms a simple refusal baseline and compares favorably to self-consistency. It also captures cases where models do not explicitly refuse. In contrast, no reliable geometric signal emerges for factual prompts, indicating that the effect is form-conditional rather than universal. Code prompts show large effect sizes with higher variance, suggesting partial generalization beyond mathematical form. A layer-wise analysis reveals that the signal arises in early layers and gradually attenuates toward the output. These results suggest that answerability-related geometry is established before the final stages of generation. Together, these findings indicate that geometric deviation can serve as a lightweight pre-generation signal that is reliable in structured domains with formal answerability constraints, with clear boundaries on where it generalizes.

Medical large language model (LLM) evaluations rely on simplified, exam-style benchmarks that rarely reflect the ambiguity of real-world medical inquiries. We introduce the CLinical Evaluation of Ambiguity and Reliability (CLEAR) framework, which assesses how decision-space presentation, ambiguity, and uncertainty affect LLMs' reasoning on medical benchmarks. CLEAR systematically perturbs (1) the number of plausible answer options, (2) the presence of a ground truth or abstention option, and (3) the semantic framing of answer options. Applying CLEAR on three benchmarks evaluated across 17 LLMs reveals three notable limitations of existing evaluation methods. First, increasing the number of plausible answers degrades a model's ability to identify the correct answer and abstain against incorrect ones. Second, this lack of caution intensifies as the framing of abstention shifts from assertive rejection like "None of the Above" to uncertainty admission like "I don't know" (IDK). Notably, just including IDK in the answer space increases incorrect answer selections. Lastly, we formalize the performance gap between identifying the correct answer and abstaining from incorrect ones as the humility deficit, which worsens with model scale. Our findings reveal limitations in standard medical benchmarks and underscore that scaling alone does not resolve LLM reliability issues.

Video moment retrieval is the task of retrieving specific segments of a video corresponding to a given text query. Recent studies have been conducted to improve multimodal alignment performance through visual-linguistic similarity learning at the snippet-level and transformer-based temporal boundary regression. However, existing models do not calculate similarity by considering the relationships between multiple answer segments that match the query. Therefore, existing models are easily influenced by visually similar segments in the surrounding context. Existing models calculate similarity at the snippet-level and ignore the relationships between multiple answer segments corresponding to a single query. Therefore, they struggle to exclude segments irrelevant to the query. To address this issues, we propose ClipTBP, a clip-pair temporal boundary prediction framework based on boundary-aware learning. ClipTBP introduces a clip-level alignment loss for explicitly learning the semantic relationship between answer segments. ClipTBP also predicts accurate temporal boundaries by applying both main boundary loss and auxiliary boundary loss. ClipTBP consistently improves performance when applied to various existing models and demonstrates more robust boundary prediction performance even in ambiguous query scenarios.

Large Audio-Language Models show consistent performance gains across speech and audio benchmarks, yet high scores may not reflect true auditory perception. If a model can answer questions without processing the acoustic signal, the benchmark fails as a measure of auditory understanding. We present a diagnostic framework using two axes: text prior, which measures answerability from text and general knowledge alone, and audio reliance, which assesses actual dependency on the acoustic signal. Evaluating eight LALMs across three benchmarks, we find that models retain 60-72% of their full audio scores even without any audio input. Moreover, among items that require audio, only 3.0-4.2% need the complete audio clip; the majority can be resolved using localized fragments. These findings challenge the assumption that benchmark performance equals robust audio understanding, and we conclude with practical guidelines for improving evaluation reliability and benchmark design.

Stochastic sampling strategies are widely adopted in large language models (LLMs) to balance output coherence and diversity. These heuristics are often inherited in Multimodal LLMs (MLLMs) without task-specific justification. However, we contend that stochastic decoding can be suboptimal for Visual Question Answering (VQA). VQA is a closed-ended task with head-heavy answer distributions where uncertainty is usually epistemic, arising from missing or ambiguous visual evidence rather than plausible continuations. In this work, we provide a theoretical formalization of the relationship between model calibration and predictive accuracy, and derive the sufficient conditions for greedy decoding optimality. Extensive experiments provide empirical evidence for the superiority of greedy decoding over stochastic sampling across multiple benchmarks. Furthermore, we propose Greedy Decoding for Reasoning Models, which outperforms both stochastic sampling and standard greedy decoding in multimodal reasoning scenarios. Overall, our results caution against naively inheriting LLMs decoding heuristics in MLLMs and demonstrate that greedy decoding can be an efficient yet strong default for VQA.

NL2SQL systems deployed in industry settings often encounter ambiguous or unanswerable queries, particularly in interactive scenarios with incomplete user clarification. Existing benchmarks typically assume a single source of ambiguity and rely on user interaction for resolution, overlooking realistic failure modes. We introduce Clarity, a framework for automatically generating an NL2SQL benchmark with multi-faceted ambiguities and diverse user behaviors across both single- and multi-turn settings. Using a constraint-driven pipeline, Clarity transforms executable SQL into ambiguous queries, augmented with grounded conversational continuations and schema-level metadata. Empirical evaluation on Spider and BIRD shows that leading NL2SQL systems, including those based on strong LLMs, suffer significant performance degradation under multi-faceted ambiguity. While these systems often detect ambiguity, they struggle to accurately localize and resolve the underlying schema-level sources. Our results highlight the need for more robust ambiguity detection and resolution in industry-grade NL2SQL systems.

Large Language Models often achieve strong performance by generating long intermediate chain-of-thought reasoning. However, it remains unclear when a model's final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using forced answer completion, which elicits the model's intermediate predictions at partial reasoning prefixes. Focusing on Qwen3-4B and averaging results across all datasets considered, we find that predicted answers change in only 32% of queries. Moreover, once the final answer switch occurs, the model generates an average of 760 additional reasoning tokens per query, accounting for a substantial fraction of the total reasoning budget. Motivated by these findings, we investigate early stopping strategies that halt generation once the answer has stabilized. We show that simple heuristics, including probe-based stopping, can reduce reasoning token usage by 500 tokens per query while incurring only a 2% drop in accuracy. Together, our results indicate that a large portion of chain-of-thought generation is redundant and can be reduced with minimal impact on performance.

We introduce a benchmark dataset for question answering and translation in bilingual Latin and English settings, containing about 7,800 question-answer pairs. The questions are drawn from Latin pedagogical sources, including exams, quizbowl-style trivia, and textbooks ranging from the 1800s to the present. After automated extraction, cleaning, and manual review, the dataset covers a diverse range of question types: knowledge- and skill-based, multihop reasoning, constrained translation, and mixed language pairs. To our knowledge, this is the first QA benchmark centered on Latin. As a case study, we evaluate three large language models -- LLaMa 3, Qwen QwQ, and OpenAI's o3-mini -- finding that all perform worse on skill-oriented questions. Although the reasoning models perform better on scansion and literary-device tasks, they offer limited improvement overall. QwQ performs slightly better on questions asked in Latin, but LLaMa3 and o3-mini are more task dependent. This dataset provides a new resource for assessing model capabilities in a specialized linguistic and cultural domain, and the creation process can be easily adapted for other languages. The dataset is available at: https://github.com/slanglab/RespondeoQA

Answering open-ended questions remains challenging for AI systems because it requires synthesis, judgment, and exploration beyond factual retrieval, and users often refine answers through multiple iterations rather than accepting a single response. Existing QA benchmarks do not explicitly support this refinement process. To address this gap, we introduce a new task, document-grounded related insight generation, where the goal is to generate additional insights from a document collection that help improve, extend, or rethink an initial answer to an open-ended question, ultimately supporting richer user interaction and a better overall question answering experience. We curate and release SCOpE-QA (Scientific Collections for Open-Ended QA), a dataset of 3,000 open-ended questions across 20 research collections. We present InsightGen, a two-stage approach that first constructs a thematic representation of the document collection using clustering, and then selects related context based on neighborhood selection from the thematic graph to generate diverse and relevant insights using LLMs. Extensive evaluation on 3,000 questions using two generation models and two evaluation settings shows that InsightGen consistently produces useful, relevant, and actionable insights, establishing a strong baseline for this new task.

Streamliner constraints reduce the search space of combinatorial problems by ruling out portions of the solution space. We adapt the StreamLLM approach, which uses Large Language Models (LLMs) to generate streamliners for Constraint Programming, to Answer Set Programming (ASP). Given an ASP encoding and a few small training instances, we prompt multiple LLMs to propose candidate constraints. Candidates that cause syntax errors, render satisfiable instances unsatisfiable, or degrade performance on all training instances are discarded. The surviving streamliners are evaluated together with the original encoding, and we report results for a virtual best encoding (VBE) that, for each instance, selects the fastest among the original encoding and its streamlined variants. On three ASP Competition benchmarks (Partner Units Problem, Sokoban, Towers of Hanoi), the VBE achieves speedups of up to 4--5x over the original encoding. Different LLMs produce semantically diverse constraints, not mere syntactic variations, indicating that the approach captures genuine problem structure.

Reinforcement Learning with Verifiable Rewards (RLVR) improves multimodal reasoning by rewarding verifiable final answers. Yet answer-correct trajectories may still rely on incomplete derivations, weak evidence, or statements that contradict their conclusions. This gap between answer correctness and reasoning validity, which we call reasoning-answer inconsistency, motivates trajectory supervision in multimodal RL. We compare two main approaches: reward models (RMs), and Generative Rewards (GRs). RMs are efficient and help early in training, but their gains weaken as the policy distribution shifts; GRs improve performance, but may give unstable rewards and computationally expensive. We therefore propose Groupwise Ranking Reward, which ranks verifier-passed trajectories for the same prompt in one pass and redistributes reward accordingly. Groupwise comparison better separates stronger and weaker correct trajectories with lower judge overhead than GRs. Experiments show that RLVR aggravates reasoning-answer inconsistency, while trajectory supervision alleviates it. Groupwise Ranking Reward performs best overall, improving reliability-conditioned accuracy from 47.4% to 54.7% over RLVR.

Developing agents capable of navigating fragmented, multi-source information remains challenging, primarily due to the scarcity of benchmarks reflecting hybrid workflows combining database querying with external APIs. To bridge this gap, we introduce ReCoQA, a large-scale benchmark of 29,270 real-estate instances featuring machine-verifiable supervision for intermediate steps, including structured intent labels, SQL queries, and API calls. Complementarily, we propose HIRE-Agent, a hierarchical framework instantiating an understand-plan-execute architecture as a strong baseline. By orchestrating a Front-end parser, a planning Supervisor, and execution Specialists, HIRE-Agent effectively integrates heterogeneous evidence. Extensive experiments demonstrate that HIRE-Agent constitutes a strong baseline and substantiates the necessity of hierarchical collaboration for complex, real-world reasoning tasks.

When language models answer open-ended problems, they implicitly make hidden decisions that shape their outputs, leaving users with uncontextualized answers rather than a working map of the problem; drawing on multiverse analysis from statistics, we build and evaluate the conceptual multiverse, an interactive system that represents conceptual decisions such as how to frame a question or what to value as a space users can transparently inspect, intervenably change, and check against principled domain reasoning; for this structure to be worth navigating rather than misleading, it must be rigorous and checkable against domain reasoning norms, so we develop a general verification framework that enforces properties of good decision structures like unambiguity and completeness calibrated by expert-level reasoning; across three domains, the conceptual multiverse helped participants develop a working map of the problem, with philosophy students rewriting essays with sharper framings and reversed theses, alignment annotators moving from surface preferences to reasoning about user intent and harm, and poets identifying compositional patterns that clarified their taste.

Agentic systems often fail not by being entirely wrong, but by being too precise: a response may be generally useful while particular claims exceed what the evidence supports. We study this failure mode as overcommitment control and introduce compositional selective specificity (CSS), a post-generation layer that decomposes an answer into claims, proposes coarser backoffs, and emits each claim at the most specific calibrated level that appears admissible. The method is designed to express uncertainty as a local semantic backoff rather than as a whole-answer refusal. Across a full LongFact run and HotpotQA pilots, calibrated CSS improves the risk-utility trade-off of fixed drafts. On the full LongFact run, it raises overcommitment-aware utility from 0.846 to 0.913 relative to the no-CSS output while achieving 0.938 specificity retention. These results suggest that claim-level specificity control is a useful uncertainty interface for agentic systems and a target for future distribution-free validity layers.

Effective abstention (EA), recognizing evidence insufficiency and refraining from answering, is critical for reliable multimodal systems. Yet existing evaluation paradigms for vision-language models (VLMs) and multi-agent systems (MAS) assume answerability, pushing models to always respond. Abstention has been studied in text-only settings but remains underexplored multimodally; current benchmarks either ignore unanswerability or rely on coarse methods that miss realistic failure modes. We introduce MM-AQA, a benchmark that constructs unanswerable instances from answerable ones via transformations along two axes: visual modality dependency and evidence sufficiency. Evaluating three frontier VLMs spanning closed and open-source models and two MAS architectures across 2079 samples, we find: (1) under standard prompting, VLMs rarely abstain; even simple confidence baselines outperform this setup, (2) MAS improves abstention but introduces an accuracy-abstention trade-off, (3) sequential designs match or exceed iterative variants, suggesting the bottleneck is miscalibration rather than reasoning depth, and (4) models abstain when image or text evidence is absent, but attempt reconciliation with degraded or contradictory evidence. Effective multimodal abstention requires abstention-aware training rather than better prompting or more agents.

Humans often specify tasks incompletely, so assistants must know when and how to ask clarifying questions. However, effective clarification remains challenging in software engineering tasks as not all missing information is equally valuable, and questions must target information users can realistically provide. We study clarification in real software engineering tasks by quantifying which types of information most affect task success and which questions elicit useful responses from simulated users. Using Shapley attribution and distributional comparisons, we identify two key properties of effective clarification: task relevance (which information predicts success) and user answerability (what users can realistically provide). We operationalize these properties as multi-stage reinforcement learning rewards to train CLARITI, an 8B-parameter clarification module, that matches GPT-5's resolution rate on underspecified issues while generating 41% fewer questions. Our results suggest that grounding reward design in empirical analysis of information impact and user answerability improves clarification efficiency.

Existing methods for quantifying predictive uncertainty in neural networks are either computationally intractable for large language models or require access to training data that is typically unavailable. We derive a lightweight alternative through two approximations: a first-order Taylor expansion that expresses uncertainty in terms of the gradient of the prediction and the parameter covariance, and an isotropy assumption on the parameter covariance. Together, these yield epistemic uncertainty as the squared gradient norm and aleatoric uncertainty as the Bernoulli variance of the point prediction, from a single forward-backward pass through an unmodified pretrained model. We justify the isotropy assumption by showing that covariance estimates built from non-training data introduce structured distortions that isotropic covariance avoids, and that theoretical results on the spectral properties of large networks support the approximation at scale. Validation against reference Markov Chain Monte Carlo estimates on synthetic problems shows strong correspondence that improves with model size. We then use the estimates to investigate when each uncertainty type carries useful signal for predicting answer correctness in question answering with large language models, revealing a benchmark-dependent divergence: the combined estimate achieves the highest mean AUROC on TruthfulQA, where questions involve genuine conflict between plausible answers, but falls to near chance on TriviaQA's factual recall, suggesting that parameter-level uncertainty captures a fundamentally different signal than self-assessment methods.

Although large language models (LLMs) show promise for generating educational feedback, it remains unclear how feedback should be designed (e.g., tone and coverage) to support answer revision and learner evaluations across learner profiles. We define six feedback designs for multiple-choice biology questions, including a baseline design and five variants with additional feedback elements, and conduct an empirical study with 321 high school students. We evaluate feedback using immediate revision performance and six subjective evaluation criteria, and analyze differences in subjective evaluations across learner profiles based on personality traits. Our results show that presenting task-relevant information clearly is associated with better immediate revision performance and is favorably evaluated across learner profiles, while we observe descriptive differences in evaluation patterns, particularly for informational novelty and affective framing. These findings support further investigation of personalized LLM feedback design.

The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.

Vision-centric retrieval for VQA requires retrieving images to supply missing visual cues and integrating them into the reasoning process. However, selecting the right images and integrating them effectively into the model's reasoning remains challenging. To address this challenge, we propose R3G, a modular Reasoning-Retrieval-Reranking framework. It first produces a brief reasoning plan that specifies the required visual cues, then adopts a two-stage strategy, with coarse retrieval followed by fine-grained reranking, to select evidence images. On MRAG-Bench, R3G improves accuracy across six MLLM backbones and nine sub-scenarios, achieving state-of-the-art overall performance. Ablations show that sufficiency-aware reranking and reasoning steps are complementary, helping the model both choose the right images and use them well. We release code and data at https://github.com/czh24/R3G.

Reinforcement Learning (RL) has become a key driver for enhancing the long chain-of-thought (CoT) reasoning capabilities of Large Language Models (LLMs). However, prevalent methods like GRPO often fail when task difficulty exceeds model capacity, leading to reward sparsity and inefficient training. Prior work attempts to mitigate this with off-policy data, but such methods often induce severe distributional mismatches that destabilize policy updates. In this work, we identify a core issue underlying these failures, which we term low training affinity, and introduce Affinity, the first quantitative metric for monitoring the compatibility between external guidance and the model's intrinsic policy. To address this, we propose HINT, an adaptive framework designed to enhance reasoning capabilities while explicitly preserving high Affinity. First, instead of revealing partial answers, HINT supplies Meta-Hints, which act as abstract cognitive scaffolding to guide the model in articulating solutions independently. Second, to ensure stability, we integrate Affinity-Aware Policy Optimization (AAPO), which dynamically modulates the learning objective based on the Affinity. Extensive experiments across diverse benchmarks demonstrate that HINT consistently outperforms strong baselines, while exhibiting superior stability and robust generalization to out-of-distribution tasks. Code is available at https://github.com/ViviqwerAsd/HINT.

Large audio-language models (LALMs) are often used in tasks that involve reasoning over ordered options. An open question is whether their predictions are influenced by the order of answer choices, which would indicate a form of position bias and undermine their reliability. In this paper, we identify and analyze this problem in LALMs. We demonstrate that no model is immune to this bias through extensive experiments on six LALMs across three widely used benchmarks and their spoken counterparts. Shuffling the order of answer options can cause performance fluctuations of up to 24% and even change model rankings, raising concerns about the reliability of current evaluation practices. We also study permutation-based strategies and show that they can mitigate bias in most cases. Our work represents the first systematic investigation of this issue in LALMs, and we hope it raises awareness and motivates further research in this direction.

Compositional visual reasoning has emerged as a key research frontier in multimodal AI, aiming to endow machines with the human-like ability to decompose visual scenes, ground intermediate concepts, and perform multi-step logical inference. While early surveys focus on monolithic vision-language models or general multimodal reasoning, a dedicated synthesis of the rapidly expanding compositional visual reasoning literature is still missing. We fill this gap with a comprehensive survey spanning 2023 to 2025 that systematically reviews 260+ papers from top venues (CVPR, ICCV, NeurIPS, ICML, ACL, etc.). We first formalize core definitions and describe why compositional approaches offer advantages in cognitive alignment, semantic fidelity, robustness, interpretability, and data efficiency. Next, we trace a five-stage paradigm shift: from prompt-enhanced language-centric pipelines, through tool-enhanced LLMs and tool-enhanced VLMs, to recently minted chain-of-thought reasoning and unified agentic VLMs, highlighting their architectural designs, strengths, and limitations. We then catalog 60+ benchmarks and corresponding metrics that probe compositional visual reasoning along dimensions such as grounding accuracy, chain-of-thought faithfulness, and high-resolution perception. Drawing on these analyses, we distill key insights, identify open challenges (e.g., limitations of LLM-based reasoning, hallucination, a bias toward deductive reasoning, scalable supervision, tool integration, and benchmark limitations), and outline future directions, including world-model integration, human-AI collaborative reasoning, and richer evaluation protocols. By offering a unified taxonomy, historical roadmap, and critical outlook, this survey aims to serve as a foundational reference and inspire the next generation of compositional visual reasoning research.