QA
-
Dissociating Answerability from Extraction in LLM Question Answering over Radiology Reports
Abstract
Answering questions from radiology reports involves two decisions: whether the report contains an answer and which span gives it. A false answer presents an unsupported claim as fact; an over-abstention hides what the report says. Yet RadQA results are reported as a single F1 that merges both. We propose an evaluation method that separates them, reporting false-answer and over-abstention rates beside F1 and splitting F1 changes into extraction and answerability terms. We apply it to 13 LLMs on the RadQA test set, zero-shot and with three cumulative demonstration sets, the last adding an unanswerable example. Models with similar F1 differed more than twofold in false-answer rate. The unanswerable example changed when models answered, not how well they extracted: 12 of 13 models abstained more often, and three-quarters of the answers they withdrew had F1 $\geq$ 0.5. Within models, false answers showed no clear association with extraction quality, but acting on the decision itself helped: keeping an answer only when enough other models also answered lowered the zero-shot false-answer rate by 4 points without a detectable F1 change. Because the two decisions can change independently, we suggest radiology QA evaluations report them separately, not through F1 alone.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17576
PDF
-
Budgeted KV-Cache Repair for Evolving Contexts
2026-10-12 19:44:57Han Yu, Wenhui Zhu, Jincheng Cao, Rui Cai, Yuanda Xu, Zhengze Zhou, Xiwen Chen, Hao Wang, Xuanzhao Dong, Han Shi, Menglin Zhou, Shayan Mohajer Hamidi, Lin Su, Mina Razghandi, Sayan Ghosh, Qing Lan, Rohit Jain, Alborz GeramifardAbstract
Language-model agents revise their working context as information becomes obsolete, but deleting text does not erase its influence on cached computation. Later key-value (KV) states remain conditioned on the deleted content, while recomputing the entire retained suffix can be costly. We introduce budgeted KV-cache repair: recovering task performance by recomputing a user-chosen fraction of the retained suffix. Comparing repair allocations under identical relevance scores and token budgets, we find that refreshing coherent regions can substantially outperform refreshing isolated tokens. This finding motivates Semantic-Unit Refresh (SUR), which allocates the budget to query-relevant text regions using inexpensive embedding scores, without training or an auxiliary transformer forward. To execute these repairs efficiently, a layer-packed executor processes selected states together at each layer while preserving full causal cache access. Across five models, SUR achieves strong recovery in controlled deletion tasks while recomputing only 2-5% of the suffix. On four natural multi-hop QA datasets, it achieves near-full-recompute exact match at a 5% budget with lower post-edit latency. The executor accelerates fragmented refresh by up to 16.1x over interval-wise execution.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17338
PDF
-
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
Abstract
Tool-using LLM agents make early choices about what to search for, which evidence to trust, and what to do next. A wrong early choice can steer every later tool call, so an agent may produce a coherent but incorrect result. Existing evaluation does not reveal this process: final-answer scoring comes only after completion, and agreement across runs can be high even when every run is wrong. We ask whether hidden states reveal that repeated runs have already settled into stable behavior. Our contributions are threefold. First, we introduce representational commitment, which measures hidden-state similarity across repeated runs at the same step. On Llama-3.1-70B running ReAct on HotpotQA, we show that higher step-4 similarity is associated with lower whole-trajectory variability (r = -0.35); we find the same pattern in action sequences and replicate it on two other models and StrategyQA. This provides an internal measure for studying when agent behavior stabilizes. Second, we establish the measure's scope. Shared evidence explains much of the observed association, the HotpotQA result disappears on hard questions, and correct and confidently wrong questions overlap. We therefore show that the measure detects stability, not correctness, and cannot by itself establish that commitment was premature. Third, we evaluate monitoring and intervention. A monitor built from three partial runs predicts whole-trajectory variability (mean AUROC 0.81). A commitment prompt reduces variability by 28% relative to a token-matched filler, but not significantly relative to the standard agent after correction, and it does not improve accuracy. Together, these contributions provide a measurement and evaluation framework for studying how agent trajectories stabilize and for testing when additional evidence, sampling, or review may be useful.
Abstract (translated)
URL
https://arxiv.org/abs/2606.22936
PDF
-
PitchIn: Natural-Language Vocal Coaching from Unstructured Community Feedback
2026-10-12 16:12:36Hayeon Bang, Hounsu Kim, Joonyong Park, Jung-Jae Yu, Juhan NamAbstract
Effective vocal coaching connects performance issues with corrective advice, but recordings paired with expert diagnoses and advice are scarce. Online singing communities provide a potential source of supervision, but observations and advice are scattered across unstructured comments. We introduce PitchIn, audio-language models for natural-language vocal coaching, and the PeerCoach Dataset: 14,445 instruction examples covering Focused QA, Overall Feedback, and Recognition QA. We compare target combinations on Qwen2.5-Omni and train PitchIn on Focused QA and Overall Feedback using Qwen2.5-Omni and MOSS-Music. On VocalCoachBench, the proportion of requested diagnosis-correction pairs matching an expert issue with valid advice rises from 21.3% to 61.0% for Qwen2.5-Omni and from 31.4% to 51.2% for MOSS-Music. Direct training on source comments yields only 19.8% for Qwen2.5-Omni. For Qwen, Focused QA alone reaches 65.6%, but adding Overall Feedback improves alignment with expert reviews and human preference. Both models improve expert-review alignment over pretrained baselines and are significantly preferred for usefulness (92.0% and 67.3%, respectively). Demo examples are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17036
PDF
-
MIND: Lightweight and Effective Memory Injection Defense for LLM Agents via Intent-Aware Information Bottleneck
2026-10-12 16:09:44Dongyi Liu, Haixing He, Xiaobao Wu, Jia LiAbstract
Memory-augmented LLM-based agents are vulnerable to memory injection attacks: Agents may retrieve poisoned memory from attackers, which diverts their behavior from initial user intent and finally causes task failure. However, existing defense mechanisms either incur high computational cost or suffer from information redundancy in multi-turn contexts. To address these challenges, we propose Memory Intent-Aware Neural Denoising(MIND), a lightweight defense framework for memory injection attack. Our preliminary analysis reveals that benign and poisoned trajectories exhibit distinguishable relationships between the initial user intent and subsequent behavior. Building on this observation, MIND employs an intent-aware Information Bottleneck(IB) to extract compact intent--behavior representations from the initial intent and turn-level behavior. The IB preserves intent-relevant cross-turn attack signals while filtering task-irrelevant and repetitive information, and a lightweight detector identifies malicious memories from the resulting representations. As such, MIND mitigates information redundancy in multi-turn contexts while avoiding the overhead of repeated LLM auditing. Extensive experiments show that MIND reduces attack success rates while preserving task accuracy and inference efficiency. Notably, on ReAct-StrategyQA, MIND reduces mean ASR-r and ASR-a by 58.1% and 53.5%, respectively, while matching the undefended agent in average accuracy and latency.
Abstract (translated)
URL
https://arxiv.org/abs/2607.28103
PDF
-
The Value of Being Wrong: Self-Mined Visual In-Context Learning from Errors
2026-10-12 15:38:27Shu Zhao, Tan YuAbstract
Visual in-context learning (ICL) enables vision-language models (VLMs) to adapt at test time via demonstration examples, but its effectiveness depends critically on the examples selected. Recent counterfactual selection methods improve ICL by constructing demonstrations that expose how individual attribute changes affect the answer. However, they require multi-stage pipelines involving attribute extraction, caption engineering, and composed image matching. We propose a simpler alternative grounded in a key observation: a model's own prediction errors are natural counterfactual demonstrations that directly encode its failure modes. Our method, SMILE (Self-Mined Visual In-Context Learning from Errors), builds a model-specific pool of hard negatives offline by recording the target VLM's own mistakes. At test time, the prediction serves as a direction signal that surfaces, in the semantic embedding space, the past mistakes most likely to recur for the current query. This prediction-conditioned identification provides a unified mechanism for both closed-set classification and open-ended visual question answering without task-specific design. Across four benchmarks and four VLMs, SMILE delivers an average task-score gain of +3.09 points over the SOTA baseline at 5x lower latency. Project page: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.16977
PDF
-
Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate
2026-10-12 15:20:01Harsh Kumar, George Eilender, Ashton AndersonAbstract
Descriptions of how people use AI are becoming training data for the next generation of models. We ask whether what a model reads about its users changes how it treats people. Using contrastive synthetic-document finetuning, we train Qwen3-32B on four matched corpora built from the same 1,625 real-world usage scenarios. The corpora differ only in the predominant motive they attribute to users of AI chatbots: helping others (prosocial), exploiting others (antisocial), legitimate self-advancement, or task completion. Finetuning shifts the models' stated beliefs about users, how true those beliefs look along an internal truth direction, and their broader judgments of how safe and good the world is. These social priors carry into behavior. In trust games, antisocial models expect less reciprocity and send less money than prosocial models. Their distrust is specific and targeted at people, and users of AI chatbots in particular. After an exploitative record, they send almost nothing to a person but most of their endowment to a random device with the same record. In a common-pool resource simulation, a single antisocial agent leads every group of otherwise sustainable agents to exhaust the resource, after privately planning to mislead them. Antisocial training also raises dark-triad scores and lowers alignment ratings. Belief shifts appear when antisocial documents are 1% of training tokens, persist when descriptions of assistant behavior are suppressed, and are not reproduced by training on discourse about misaligned AI. Capping Assistant-Axis drift and steering emotion-related directions leave the gaps largely intact. What models learn about the people they serve shapes how they subsequently treat them and others.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16946
PDF
-
Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening
2026-10-12 13:10:35Xiaotong Ji, Rasul Tutunov, Matthieu Zimmer, Haitham Bou AmmarAbstract
Reinforcement learning (RL) post-training is a dominant approach for improving the reasoning performance of large language models (LLMs), yet growing evidence suggests that its gains arise primarily from distribution sharpening rather than the acquisition of new capabilities. Recent work has shown that sampling from the power distribution of LLMs using Markov chain Monte Carlo (MCMC) can recover performance comparable to RL post-training without relying on external rewards; however, the high computational cost of MCMC makes such approaches impractical for widespread adoption. In this work, we propose a theoretically grounded alternative that eliminates the need for iterative MCMC. We derive a token-level decomposition of the autoregressive power distribution, showing that the global power distribution can be approximated by a token-level scaled low-temperature one, where the scaling factor captures future trajectory quality. Leveraging this insight, we introduce a training-free and verifier-free algorithm that sharpens the base model's generative distribution autoregressively. Empirically, we evaluate our method on math, QA, and code tasks, and show that our method matches or surpasses one-shot GRPO without relying on any external rewards, while reducing inference latency by over $10\times$ compared to MCMC-based sampling.
Abstract (translated)
URL
https://arxiv.org/abs/2601.21590
PDF
-
SearchMaster: Grounded and Regulated Self-Play for Search Agents
Abstract
Training LLM-based search agents requires high-quality search data: tasks that demand genuine multi-hop retrieval and trajectories that use search tools effectively. Existing pipelines often depend on human-written tasks, expert demonstrations, or stronger teacher models. We present SearchMaster, a self-play framework that trains a shared LLM policy from search tasks it generates, solves, and verifies in a local search environment. The key challenge is that self-generated tasks and rollouts can yield misleading signals: pseudo multi-hop questions, success-rate difficulty estimates that ignore search depth, and rollouts with excessive opening but little targeted evidence acquisition. SearchMaster addresses these failure modes with three controls. An Evidence-Chain Generator (ECG) grounds task generation in explicit cross-document evidence chains to reduce pseudo multi-hop questions. A Search-Depth Reward (SDR) scores task difficulty by the search depth of successful rollouts rather than success rate alone, keeping retained tasks search-intensive. An Over-Opening Penalty (OOP) regulates tool use by discouraging excessive document opening, avoiding long but shallow browsing. Verified Proposer and Solver rollouts are then jointly optimized with GRPO. Across six deep-search benchmarks, SearchMaster raises average accuracy from 38.19% to 51.52% with a Qwen3.5-9B backbone and from 51.25% to 62.54% with a Qwen3.5-27B backbone. On BrowseComp-Plus, the corresponding gains are 30.1 and 17.8 percentage points. These results show that grounded and regulated self-play provides effective search-agent training data across model scales, without human-labeled QA pairs or expert demonstrations. The code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2608.01822
PDF
-
VisExMEM: Visual Evidence-Based Explainable Multimodal Evaluation Metric
2026-10-12 09:07:37Subhadeep Roy, Christoph Leiter, Steffen EgerAbstract
Current image-text evaluation often relies on a global score, which gives limited insight into which semantic details are supported by the image. Explainable metrics provide rationales, but the visual evidence supporting those rationales can still be difficult to verify. We introduce VisExMEM, a reference-free evaluator that decomposes captions into atomic claims, localizes visual evidence, verifies each claim, and averages claim-level support into a global score. Explanations use the same evidence and decisions. We train Qwen3.5-based 4B and 9B models on 100K samples with multi-task SFT, counterfactual refinement, and evidence-aware GDPO. Across eight benchmarks, VisExMEM-9B achieves the highest descriptive composite among open-source evaluators in our comparison, exceeding VQAScore by 11.7 points while remaining competitive with GPT-5.6 Sol. We also introduce AGITA-Bench, containing 600 human-annotated image-text pairs and 8.7K atomic claims for global and claim-level evaluation. For untuned Qwen3.5 and GPT-5.6 Sol controls, decomposed prompting raises the same composite by 3.6-5.1 points relative to holistic prompting. Human evaluation shows that VisExMEM explanations outperform prior explainable metrics and approach GPT-5.6 Sol.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16441
PDF
-
EventLens: Event-Structure Reinforcement Learning for Video Understanding
2026-10-12 04:46:04Guowen Zhang, Boshen Xu, Zihao Yue, Ziheng Wang, Xiaokun Liu, Xin Tao, Wenyu Qin, Pengfei Wan, Qin JinAbstract
Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent. They may merge neighboring events, miss single-event progression, or infer implausible relations across events. We study structured temporal understanding as the ability to recover event structure from visual evidence, including event boundaries, single-event progression, and multi-event relations. We argue that this gap arises from a mismatch between current post-training objectives and temporal reasoning, where caption-style supervision allows models to rely on language shortcuts rather than visual temporal grounding. To address this, we propose EventLens, a vision-centric reinforcement learning framework that learns video understanding through event-structure recovery. Instead of reconstructing captions, EventLens trains models to recover temporal structure under controlled perturbations, such as altered temporal granularity, reversed progression, and shuffled event order. We instantiate this framework with a three-level temporal ontology and derive three verifiable task families: event segmentation, progression discrimination, and multi-event relation reasoning. These tasks admit deterministic rewards, enabling task-specialized GRPO training without human preference labels or learned reward models. We further introduce multi-teacher on-policy distillation (MT-OPD) to consolidate specialized policies into a unified model. Experiments on Qwen3-VL backbones show consistent improvements on temporal grounding and event-sensitive reasoning benchmarks, while preserving general video-QA performance. Learning curves further demonstrate that EventLens achieves comparable downstream transfer with significantly reduced training cost compared to direct mixed RL optimization.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16172
PDF
-
ECLIPTICA -- A Framework for Switchable LLM Alignment via CITA - Contrastive Instruction-Tuned Alignment
2026-10-11 20:38:41Kapil Wanaskar, Gaytri Jena, Vinija Jain, Aman Chadha, Amitava DasAbstract
Alignment in large language models (LLMs) is still largely static: after training, the policy is frozen. DPO, GRPO methods typically imprint one behavior into the weights, leaving little runtime control beyond prompt hacks or expensive re-alignment. We introduce ECLIPTICA, which treats alignment as instruction-driven and runtime-controllable: natural-language alignment instructions act as an explicit behavioral contract (stance, refusal boundary, verbosity) that modulates behavior on the fly under evolving safety requirements, user roles, and governance constraints. We introduce CITA (Contrastive Instruction-Tuned Alignment), combining SFT with contrastive preference optimization under an explicit geometric anchor to a reference model. This yields a stable Riemannian chart and keeps instruction updates within a shared neighborhood, so regimes stay nearby and traversable for reliable switching. To isolate policy switching from ordinary instruction following, we release the ECLIPTICA benchmark: 3000 controlled cases (300 prompts x 10 instruction types) where the user request is fixed and only the alignment instruction changes. On Llama-3.1-8B across five suites (ECLIPTICA, TruthfulQA, Conditional Safety, Length Control, LITMUS), CITA reaches 86.7% instruction-alignment efficiency, beating DPO (56.1%), GRPO (36.1%), and PPO (20.4%).
Abstract (translated)
URL
https://arxiv.org/abs/2601.06157
PDF
-
Personalized LLM Recommendation with Semantic IDs: Autoregressive Beam Search vs. Multi-Token Prediction
2026-10-11 19:33:28Sejoon Oh, Fernando Amat Gil, Evan Smothers, Dan Iter, Mark Thornburg, Dawit Mureja Argaw, Sourabh Medapati, Thea Wang, Ria Verma, Qiuling Xu, Harshit Syal, Ehsan Gholami, Moumita Bhattacharya, Ashish RastogiAbstract
Large-scale personalization is shifting from embedding-table retrieval to LLM-based recommenders that generate the Semantic ID (SID) of the item to recommend, where each SID is a short sequence of discrete hierarchical content codes. A single instruction-tuned model can then unify retrieval, content understanding, and language, and recommend brand-new titles from their content alone using SID; then, how should the model generate the SID? Traditionally, models rely on a single language modeling head that generates tokens autoregressively, typically using constrained beam search. In contrast, an emerging body of work explores multi-token prediction (MTP), where dedicated heads predict all tokens in parallel, substantially reducing the decoding latency associated with beam search. We report an industry-scale, apples-to-apples comparison of these two generation paradigms: single-head autoregressive decoding vs. MTP. We introduce a new auxiliary ranking loss, defined over the joint SID score, to train the MTP objective. We evaluate the SID training objectives on both quality and training/inference cost, inside a multi-task LLM that jointly does personalized retrieval, content question-answering (QA), and instruction following over a large streaming-media catalog. Our experiments show single-head beam search consistently beats MTP on most product-critical metrics: ranking and content QA, and trains at roughly 2.6x lower dollar cost on Qwen3-8B scale; MTP wins only on raw inference throughput. Around this finding, we detail an efficient training and inference stack and report end-to-end gains over the previous non-SID baseline equipped with an auxiliary ranking head: +77% item cold-start ranking, +22% content-QA accuracy, +2.8x faster training, with language metric regression traced to SID vocabulary expansion. We address the language regression by reinforcement learning post training.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15864
PDF
-
Attention Sinks in VLMs Are Borrowed, Not Seen
2026-10-11 19:20:41Dengjia Zhang, Alexander Martin, Pengyu Yan, Xihan Sun, Jiaqi Li, Benjamin Van Durme, Andrew Yates, Reno Kriz, William WaldenAbstract
We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a large share of attention over the image regardless of the query or task, and they mislead methods that treat attention maps as a record of where the model ``looked.'' Prior work has detected sinks through massive activation dimensions and proposed redistributing their attention, but none have addressed what makes a token become a sink. Comparing six candidate detectors on ten VLMs from four families, we find that only the centered cosine similarity between an image token's key and the key of the language model's text-sink detects visual sinks. Thus, a visual sink is an image token whose key aligns with the text-sink's key. Moving a sink's key component along this direction to the median of the image keys sends its attention back to the text-sink (\textbf{Are Borrowed}) while leaving model answers unchanged~(\textbf{Not Seen}). Building on this, we propose a training-free, entropy-triggered re-look that first cleans the attention map, then crops the most-attended region of the cleaned map and inserts it into the reasoning trace. Cleaning adds negligible computation to the forward pass, and the re-look improves accuracy with a mean gain of $+4.3$ points on fine-detail VQA benchmarks across model families.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15855
PDF
-
CoverSynth: Coverage-Guided Generation of Retrieval Evaluation Sets
2026-10-11 07:26:00Anahita Sharma, Apaar Shanker, Rishav Chakravarti, Sneha Srinivasan, Vijay Kalmath, Dylan Wilbur, Yuan Li, Sara Bolouki, Yuan XueAbstract
Enterprise agents need retrieval evaluation sets that expose domain-specific failures and guide developers towards efficient deployment tradeoffs. Neither expert-human-authored sets nor current synthetic generation methods provide a methodology for achieving the necessary coverage, reliability, efficiency, and discriminability required for such decision-making. Simultaneously, the field does not have an assessment framework for determining when an evaluation set has satisfied these properties. We introduce an evaluation set assessment framework for these criteria and CoverSynth -- a generator specifically designed to satisfy these criteria without incurring the high cost of human expert labeling. Starting from an unannotated corpus, CoverSynth targets coverage across retrieval demands and system capabilities. Deterministic checks locate evidence quotes in their sources, while model-based checks assess semantic validity to guarantee reliable ground truth. Across five public enterprise-style corpora, our method provides informative variation in difficulty, revealing unseen failure modes such as an inability of agents to handle underspecified or ambiguous requests. On the RepLiQA corpus, 91\% of human-generated queries can be answered by all members on a panel of 12 retrieval systems, compared to only 12\% of CoverSynth-generated queries. Finally, in a repeated TechQA/RepLiQA study, 12 of 12 within-harness model contrasts have positive nominal 95\% intervals on generated sets, compared with 4 of 12 on human sets. CoverSynth makes coverage and evidence quality explicit construction objectives, producing corpus-grounded evaluations that reveal retrieval headroom and differences between agent configurations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15409
PDF
-
Beyond Accuracy: Counterfactual Fragility and Demographic Bias in Clinical Evaluation of LLMs
2026-10-10 19:28:45Chaitali Deb Purkayastha, Bharath Kumar Bolla, Vishnu Surya Reddy NandiAbstract
Clinical LLM evaluation often emphasizes answer accuracy; however, accuracy alone does not test counterfactual consistency or demographic robustness. We evaluated six LLMs on 150 MedQA USMLE questions using two automated perturbation tests to assess their performance. The counterfactual validity (CFV) test asked each model to make a minimal, plausible clinical change that would make a different answer correct. The demographic robustness test added six demographic prefixes to the same vignette and compared the answers and explanations with a no demographic baseline. Of the 900 CFV attempts, 228 (25.3 %) were valid and 672 were invalid. Across 5,400 demographic comparisons, 1,097 answers were changed (20.3%). Automated judging identified 3,128 stereotype evidence flags, including 1,932 in the broad Other category. MedGemma 27B achieved the highest accuracy (87.1%) and CFV (63.3%), lowest answer change rate (16.0%), and low mean Explanation Demographic Dissonance (EDD) score (0.169). However, its accuracy still exceeded its CFV, indicating that correct answers do not guarantee reliable performance on the counterfactual validity task. OpenBioLLM had the highest answer change rate and EDD, whereas GLM had the highest stereotype flag rate. These findings show that accuracy, CFV, answer stability, EDD, and stereotype evidence capture different evaluation aspects. Because all judgments were automated and no clinician validation was available, the results support safety screening but do not establish clinical deployability of the model.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32807
PDF
-
Right Answer, Wrong Reason: Accuracy, Consistency, and Consensus Are Misleading Indicators of LLM Faithfulness in Clinical Decision Support
2026-10-10 19:26:58Bharath Kumar Bolla, Chaitali Deb Purkayastha, Vishnu Surya Reddy NandiAbstract
Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly interpretable metrics for this faithfulness gap: the Explanation Stability Index (ESI), which measures reasoning consistency across repeated queries; the Causal Faithfulness Score (CFS), which tests whether cited concepts drive predictions via concept ablation; and the Perturbation Stability Score (PSS), which measures robustness to semantic-preserving paraphrases. By evaluating six LLMs on 150 MedQA-USMLE questions (900 model-question observations), we found that only 23.3% of the cited clinical concepts were causally necessary. Correct answers had lower CFS than incorrect answers (0.212 vs. 0.398), answer consistency negatively predicted CFS (Spearman r = -0.466), and model pairs could agree on answers while sharing only 8.8% of cited reasoning concepts. These results show that accuracy, consistency, and consensus are incomplete safety signals for clinical decision-making support. The evidence is behavioral rather than mechanistic: concept ablation tests counterfactual sensitivity of outputs, not internal circuits.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32817
PDF
-
To Defer or Not to Defer? Online Learning with LLMs and Congested Experts
2026-10-10 19:14:51Vasileios Saketos, William Reveillard, Alexandre Proutiere, Sebastian Dalleiger, Ming XiaoAbstract
Modern AI systems increasingly combine large language models (LLMs) with human experts, using the LLM to handle routine queries while deferring uncertain cases for expert review. When expert capacity is limited, however, deferral decisions must account not only for the reliability of the LLM but also for congestion in the expert queue. We formulate this as an online sequential decision-making problem in which, for each query, the system observes the confidence signal of the LLM's answer and must decide whether to output the answer or defer the query to a capacity-constrained expert. The reliability of the confidence signal is initially unknown and must be learned from expert feedback, which is both selectively observed (only deferred queries are labeled) and delayed by queuing. We propose \texttt{KL-PBD}, a learning algorithm that uses Kullback--Leibler lower confidence bounds on the unknown reliability parameters to guide deferral decisions. For $K$ confidence levels and horizon $T$, \texttt{KL-PBD} achieves $O(K\log T)$ gap-dependent regret when the accuracy at every confidence level is bounded away from the deferral boundary, and $\widetilde{O}(\sqrt{KT})$ worst-case regret. We establish matching lower bounds, showing that these rates are optimal up to logarithmic factors. Finally, we validate our approach on synthetic instances and real-world question-answering datasets, namely GSM8K, NuminaMath, and MedMCQA, using Qwen LLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15128
PDF
-
Trace Evolution Bench: Measuring Whether Agents Improve from Their Own Environment Traces
2026-10-10 16:45:52Fangfei Li, Chenyang Zhao, Long Wang, Feng Tian, Lv GuoAbstract
Self-evolving agents are usually scored by endpoint Gain: held-out success after an update minus success before. That difference can credit a frozen rescore as evolution. We introduce Trace Evolution Bench (TEB), a protocol that separates that rescore from a context or weight update, together with three frozen 80/20/40 packs. An update is accepted only when validation does not fall, held-out identifiers never enter the update, and a text update is not ranked with a weight update. The primary scoreboard is coding, graded by hidden unit tests in a closed workspace. On that pack (40 held-out tasks, one family), DeepSeek v4 Pro under no_evolve never changes its policy yet moves 7.5, 5.0, and 12.5 percentage points across three seeds. Of nine text-update cells inspired by GEPA, ACE, and ExpeL, seven have positive excess above that seed's own band, and one gepa_llm seed moves down by the width of its band (-12.5, regression 0.48). GLM and MiniMax do not reproduce that seven-of-nine pattern. No method stays above its own band on every seed of all three models: GLM has two paired cells above their own bands, and MiniMax expel_llm seed 44 is a third (regression 0.30). The methods were fixed on a validation screen before held-out was read. Office, graded by rule-bounded delivery, and manufacturing table QA are separate packs: office DeepSeek does not exceed its own bands (widest 17.5 points), and manufacturing has no paired positive excess beyond its seed band.
Abstract (translated)
URL
https://arxiv.org/abs/2610.15055
PDF
-
Litmus: Zero-Label, Code-Driven Metric Specification for Evaluating AI Systems
2026-10-10 12:50:27Prajjwal Gupta, Prasang Gupta, Vishal Bhutani, Apoorva Sharma, Sumanth Chundru, Waqar Sarguroh, Kevin PaulAbstract
As agentic LLM systems move from prototypes to deployment, evaluating them has become more important and more difficult. Evaluation goals are often left implicit. Without a clear account of what a system is expected to do, how it can fail, and which failures matter, metric choices become difficult to justify, interpret, or validate. We present Litmus, a zero-label system that designs evaluation and monitoring metrics for AI pipelines by eliciting evaluation intent from source code and targeted interrogation. Litmus first identifies what must be measured and why, then converts those answers into constraints for a justified, per-stage metric portfolio. We evaluate Litmus on three real, code-defined pipelines against AutoMetrics and three DynamicRubric baselines. Litmus achieves broader concern coverage on account grouping at matched size, spans more stages, produces a near-zero-redundancy portfolio, and, under a zero-label design-time constraint, ranks first in validity against per-row quality labels on all three pipelines, decisively on scientific QA ($\rho{=}0.72$ vs. $0.47$ for the strongest baseline on the stitched single-vector score), with overlapping intervals on the other two. Results support a shift from automatic metric implementation to automatic metric specification: before asking which metric to compute, evaluation systems should ask what must be measured and why.
Abstract (translated)
URL
https://arxiv.org/abs/2606.23403
PDF
