Chat
-
Billiger.de Products: A Bilingual Entity Matching Benchmark
2026-09-29 14:44:02Aaron Steiner, Ksenia Elagin, Ralph Peeters, Johannes Knopp, Christian BizerAbstract
Existing product matching benchmarks primarily contain English-language product data and are often dominated by a single product category, such as electronics. This paper introduces this http URL Products, a bilingual German and English entity matching benchmark covering thirteen consumer product categories, including difficult-to-handle categories such as clothing and furniture. The benchmark data originates from the German price comparison platform this http URL. Following the design of WDC Products, the benchmark offers multiple variants that differ in the fraction of corner cases, the size of the development set, and the fraction of entities unseen during training. An aligned English translation of every offer keeps all pairs, splits, and labels fixed, while cross-language test sets combine German and English records within individual pairs. We validate the benchmark using six supervised matchers and zero-shot GPT-5.2 on both language versions and the cross-language test sets. The validation shows the difficulty of the benchmark. The comparison of the results on the English version of the benchmark to the results on the German version shows that most matchers score on average higher on the English version. The difference is largest for RoBERTa and HierGAT, while the zero-shot LLM runs are largely insensitive to the language. Comparing the F1 scores achieved by PLM-based matchers on the English version of this http URL Products with their performance on existing English-language benchmarks, such as WDC Products and Abt-Buy, shows that this http URL Products is more difficult than these benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37713
PDF
-
Correct, Don't Delete: Mitigating Emergent Misalignment with Corrective Supervision
Abstract
Fine-tuning a language model on a narrow set of harmful demonstrations, such as bad medical advice, can make it broadly misaligned on unrelated questions, a phenomenon known as emergent misalignment (EM). The usual defense is to find the offending rows and delete them, but a row locator failed our held-out test and deleting rows helps less than expected. We ask a different question: given a fixed set of poisoned rows, is it better to correct them than to remove them? We fine-tune Qwen2.5-14B-Instruct on a mixture of bad medical advice and benign chat data, select a quarter of the poison rows in advance, and either delete them or replace each with a corrected answer to the same prompt, keeping everything else the same. Replacing the rows cuts the EM rate by about a third and improves answers on held-out medical questions, while deleting the same rows has little measurable effect. The advantage is larger when half the poison rows are corrected, and it holds on a second base model and a second misaligned model organism. The content of the replacement appears to matter: paraphrasing the rows while keeping their bad advice shows no clear benefit, and the correct answers distributed with the dataset appear to do about as well as our rewriter's. Realigning an already-poisoned model with further fine-tuning is known to work, but which data does the work has not been compared directly. We find that a short round of training on corrections beats the same amount of training on generic chat data, that corrections on other medical prompts do roughly as well as corrections of the poisoned prompts themselves, and that instructing the correction writer to model a careful, harm-avoiding assistant adds no measurable benefit over plain corrections. In the settings we tested, correcting harmful training data reduces EM more than deleting it.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37624
PDF
-
Authority Bias in Language Models: Source Deference and User Agreement Are Not Interchangeable
2026-09-29 13:58:04Abhinav Rajeev Kumar (Lossfunk), Paras Chopra (Lossfunk)Abstract
Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on their merits rather than deferring to the user. Yet the same models are far more compliant when a wrong answer is attributed to a verified source, which is how retrieval results, tool outputs, and grounded-search content often present information. We measure this gap across five open-weight families and three closed APIs. A single verified-source note endorsing a wrong answer flips 45-88% of baseline-correct responses in seven of eight models, and compliance rises with how authoritative the note sounds. Source deference and user agreement are not behaviorally interchangeable inside the model: on matched items with the same wrong answer, causal interventions can selectively suppress one without equally affecting the other. In three open-weight families, removing a fitted source direction lowers source compliance by 65-80 percentage points while removing a user or assistant direction has far smaller effects, and removing the user direction shows the reverse preference. A separately fitted intervention derived from source-versus-user cue activations moves compliance in both directions while leaving the prompt text unchanged. An authority direction fitted on trivia also transfers to PIQA and multi-turn SYCON dialogues without refitting, and removing it lowers wrong-source compliance by tens of percentage points in four of five families with no detected change in MMLU-Pro or GSM8K accuracy at our evaluation sizes. Source deference and user agreement therefore need separate evaluation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37616
PDF
-
FOCUS: Training-Free Decision-Preserving Context Compression for LLM Agents
2026-09-29 13:45:34Shantanu Dixit, Anson Bastos, Xuchao Zhang, Chetan Bansal, Saravan RajmohanAbstract
LLM agents accumulate interaction histories that grow linearly with task length, causing quadratic inference cost scaling and performance degradation from attention dilution. Existing context-compression methods learn what to discard offline: by contrastively optimizing guidelines, distilling compressors, or training compression policies. This incurs a substantial cost. Further, the compression policy is learned a priori and is not dynamically conditioned on the evolving test-time trajectories. In this paper we ask a complementary question: Which past interactions causally shape the agent's future decisions? We recast context compression as a causal decision preservation problem over discrete interaction units and introduce FOCUS, a training-free context compression framework that operates entirely at test time. Our method requires no offline data collection or fine-tuning, and is architecture-agnostic, attaching to any closed-API frontier model as a modular compression layer. We evaluate FOCUS on diverse agentic benchmarks including API and tool-calling, QA, web domain and multi-turn dialogue. Our method establishes new state of the art performance, cutting peak context by up to 48% and dependency by 73% while improving task success by up to 8.9 percentage points over uncompressed execution.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37590
PDF
-
RoboHarn-Evo: Evolving Hierarchical Physical Knowledge for Self-Improving Robotic Manipulation
2026-09-29 13:42:22Shifeng Bao, Fanding Huang, Yihan Lin, Youhe Feng, Guanlin Li, Chen Zhao, Yang Li, Jiawei He, Cheng Chi, Jing ZhangAbstract
Vision-language models can coordinate long-horizon robot manipulation, yet successful task reasoning still depends on whether local physical interactions produce the intended effects. We study how repeated interaction can improve this capability without updating the base model. We introduce RoboHarn-Evo, a dual-loop harness that evolves Hierarchical Physical Knowledge (HPK) from physical experience. HPK couples two levels of reusable knowledge: Task Knowledge captures which subtask should be executed and when it is complete, while Action Knowledge captures object-relative geometric strategies and their physical effects. During execution, the agent retrieves knowledge at the corresponding decision level and grounds it in the current scene under the task goal. Across episodes, physical feedback is used to revise historical knowledge, update its applicability, and organize reusable entries for subsequent retrieval. Experiments on RMBench show that HPK improves average success by up to 24.2 percentage points across different agent models. With 80 interaction rollouts, held-out success rises from 48.3% to 75.0% for GPT-5.5 and from 70.0% to 88.3% for GPT-6. RoboHarn-Evo also resolves over 83% of historical knowledge errors while retaining 95.8% of valid knowledge, and transfers zero-shot from RMBench to RoboDojo with gains of 35.0 and 25.0 percentage points. These results demonstrate that physical interaction can be accumulated into reusable knowledge for improving subsequent manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37583
PDF
-
Quantifying Behavioral Tails in Black-Box Language Models
2026-09-29 13:41:26Elsayed Eshra, Ali Al-Lawati, Dongwon Lee, Suhang WangAbstract
We introduce RareTrap, a framework for estimating the probability of severe behaviors in black box large language models (LLMs). A key challenge for probability estimation is defining a tractable distribution over the input space. To accomplish that, RareTrap uses a surrogate LLM and constructs a geometry-aware mapping from a lower-dimensional latent reference space into its token-embedding space to induce an explicit and reproducible distribution over input prompts. A response-level performance function is utilized on the response to quantify behavior severity. This enables sequential rare event simulation that concentrates evaluations on progressively more severe behaviors while preserving probability under the induced prompt distribution, which would otherwise be prohibitive to measure. Across 10 open-weight and two frontier models (GPT-5.4 and Claude Sonnet 4.6), we find that RareTrap successfully induces severe resource consumption behaviors and computes their probability with as few as 200 evaluations. RareTrap provides model developers a principled approach for evaluating language models under a common distribution, and prioritizing alignment effort to improve safety and mitigate risks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.33638
PDF
-
Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
2026-09-29 13:10:14Merve Atasever, Keyan Azbijari, Cagan Bakirci, Bo-Ruei Huang, Tolga Izdas, Zahra Shahrooei, Richard Yang, Erdem Biyik, Jyotirmoy V. DeshmukhAbstract
Video-based policy learning is particularly promising, as it illustrates target behaviors without requiring action annotations or embodiment-matched demonstrations. A central challenge is deciding what information should be transferred from the video to the robot. Existing approaches commonly convert visual observations into scalar similarity or value signals, or ask foundation models to directly generate reward code. These approaches can make the temporal structure of a task difficult to inspect, ground, and reuse. We present Video2STL, a framework that converts observation-only videos into parametric Signal Temporal Logic (STL) specifications and uses the resulting formal representation for robot learning. A vision-language model extracts an embodiment-independent semantic event trace and constructs a bank of symbolic temporal specifications. The model determines the task structure, while numerical predicate thresholds and temporal bounds are grounded from successful robot trajectories. For policy learning, we separate short- and long-timescale temporal information: short-horizon specifications provide dense rewards through rolling-window quantitative robustness, while a causal monitor over a retained long-horizon specification provides one-time progress rewards for valid temporal prefixes. The same representation supports cross-embodiment transfer from human or animal videos to robot control. Across four manipulation tasks, Video2STL achieves $85.8\%$ average success-once and $67.0\%$ success-at-end, compared with $81.5\%/59.5\%$ for native dense PPO and $65.0\%/42.3\%$ for Text2Reward; in quadruped locomotion, Qwen-3.8 and GPT-5.6-based Video2STL policies achieve $100\%$ success across velocities from $0.3$ to $2.1\,\mathrm{m/s}$ while remaining competitive in high-speed energy efficiency. Project webpage: \href{this https URL}{video2stl}.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37519
PDF
-
Learning to Retrieve Missing Evidence for Long-Term Memory QA
2026-09-29 13:02:16Yi-Xuan Deng, Yi Zhang, Wei Liu, Chao Xue, Shuojin YangAbstract
Long-term memory enables language models to use past interactions in future conversations. However, evidence needed to answer a question may be scattered across distant turns, while the question itself omits clues needed to locate it. Retrieved facts can reveal these clues, motivating retrieval decisions conditioned on evidence already found. We introduce MERA (Missing-Evidence Retrieval Augmentation), which separates globally searchable memory from a question-specific evidence state. Verified evidence guides subsequent retrieval without restricting access to the global memory. We train a lightweight planner through reinforcement learning, rewarding queries that recover previously missing evidence. MERA achieves strong answer accuracy across Qwen3-30B and GPT-4o-mini backbones. With Qwen3-30B for evidence processing and answer generation, the trained 0.6B planner achieves 77.40% accuracy on LoCoMo and 71.29% on LongMemEval-S, exceeding a 30B planner without retrieval-grounded training by 4.10% and 3.96%, respectively. On LoCoMo, later retrieval rounds increase cumulative evidence recall from 55.5% to 80.5%.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37443
PDF
-
LazySloth: Bounded LLM-based Lazy Tree Search for Fast Long Video Comprehension
2026-09-29 12:53:57Arka Mukherjee, Kaleen Shrestha, Larissa Zhu, Maja Matari\'cAbstract
Modern vision-language models (VLMs) have shown promising results in long-video understanding due to the rich semantic information they can capture. However, most methods focus on coarse captioning of extracted image frames that are computationally inefficient and require models with large context windows. While past work has explored efficient methods through multimodal retrieval-augmented generation (RAG), they rely on lossy embeddings that lose temporal context and fine-grained detail. Few works to date have investigated how VLM-based query-relevant information retrieval can be optimized. We introduce LazySloth, an efficient tree-based search method that speeds up video comprehension and retrieval tasks 2.9-8.3x (compared to existing agentic methods) through bounded captioning of portions of the video considered irrelevant by a VLM of the video. Compared to contemporary specialized video-understanding VLMs and RAG-based methods, LazySloth achieved similar or better final task accuracy across two recent open-source base VLMs--Gemma 4 31B and Qwen3.6 27B--across four benchmarks. LazySloth reduced the gap between the base open-source model and a closed-source model, GPT-4o. Ablations showed that replacing VLM scene understanding with CLIP-based retrieval cost 8.8-19.9% in accuracy, while lazy tree construction matches eager construction at a fraction of the captioning cost. With LazySloth, we demonstrate the possibility of faster long-video comprehension without substantial loss in performance.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37426
PDF
-
Knowing Is Not Choosing: What Explicit Verification Adds Beyond Generative Preference
2026-09-29 12:05:15Yilong Li, Chengpo Yan, Aayan Arish, Suman BanerjeeAbstract
Generating a correct answer does not mean that a language model will select it. We separate factual recall into three steps: generating a correct candidate, ranking the available candidates, and selecting the final answer. Pre-generation readouts predict factual recall and which questions sampling will cover across three model families, but say little about whether an available correct answer will ultimately be selected. Explicit verification with $P(\mathrm{True})$ improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, with AUROC gains of $0.08$--$0.12$. In a prospectively defined Gemma cohort, verification raises plurality accuracy by about $5$ points, and still gains about $2$ points over chat-template likelihood, a stronger generative baseline. The advantage is strongest for relations with common-answer priors and depends on access to the entity; masking the entity removes the ranking advantage in larger Qwen models. Finally, the measured benefit depends on how correctness is defined: recall-oriented reference matching can credit option lists favored by likelihood and substantially understate the improvement seen under human semantic judgments. Prior work shows that models can carry latent factual knowledge and judge candidate answers; we show that these capabilities do not collapse into a single notion of ``knowing,'' and trace where information is gained, lost, or mismeasured between availability, ranking, and final choice.
Abstract (translated)
URL
https://arxiv.org/abs/2609.33142
PDF
-
When Tools Silently Lie: Evaluating and Mitigating Blind Compliance in Tool-Augmented Data Agents
Abstract
Tool-augmented data agents rely on tool outputs for analytical decisions. Yet successful execution can return plausible but incorrect evidence, requiring agents to decide whether to trust or verify it. Understanding this failure requires examining both the evidence obtained through checking and the answer ultimately adopted. We introduce ToxicBench to measure checking and adoption under numerical, label, schema, and retrieval errors, pairing clean and poisoned observations over fixed source data. In the 118-task GPT evaluation across three adapters, poisoning lowers task success by 26 to 39 percentage points. Ordinary retries help under one-shot poisoning, whereas repeated poisoning reveals wrong-answer adoption after checking. Controls on three public tables isolate how supplied evidence affects recovery. After freezing the scorer, we compare its judgments with human annotations on 200 trajectories, finding 96% task-success agreement. Human judgments support retry gains over Base and confirm adoption after checking on audited tasks. We release trajectories, versioned scoring, and reference and delivery audits. These findings highlight evidence availability and answer selection as complementary dimensions of agent reliability.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37153
PDF
-
Multimodal Detection of Higher-Order Behavioral Constructs: Self-Compassion in Structured Reflective Interaction
2026-09-29 09:42:22Siddhant Jain, Dimitra TsovaltziAbstract
Many of the qualities that matter most in how people learn and grow, how someone regulates their emotions, reflects on a setback, or stays aware of others during a difficult conversation, are not directly observable. They have to be inferred from how someone speaks, moves, and sounds over time, and they resist the kind of clean labeling that most machine learning pipelines are built around. We study this challenge through a case that is well grounded in psychological theory but rarely modeled computationally: self-compassion, the tendency to respond to one's own setbacks with patience rather than harsh self-criticism. We examine how it appears during structured reflective interviews in a technology-mediated training setting, where people naturally talk through socio-emotionally demanding situations. Since no existing dataset captures this kind of construct in this kind of setting, we collected and annotated 51 reflective dialog sessions using an independent, temporally overlapping annotation scheme grounded in established theory. We consolidate the underlying six-component psychological model into a three-class supervision space, balancing self-kindness and mindfulness against self-critical or overwhelmed states, and build a reproducible window-based pipeline that aligns video, audio, and text on a shared timeline. Unimodal models trained on each modality separately are compared against a simple probability-level fusion strategy, which yields modest but consistent gains over the best single modality. We close by discussing where each modality succeeds or struggles, what this suggests about how this kind of construct is actually expressed in reflective speech, and what would be needed to model it, and constructs like it, more effectively.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37148
PDF
-
Real2Gym: Building Gyms from Videos, Bringing Skills to Robots
2026-09-29 09:16:49Kerui Ren, Yingxiang Xu, Kaiwen Song, Linning Xu, Bo Dai, Mulin Yu, Tao LuAbstract
Real-world videos provide rich demonstrations of manipulation, but turning them into reusable robot skills requires visually aligned environments, executable physical interactions, and mechanisms for learning from experience. We introduce Real2Gym, an agentic Real2Sim2Real framework that turns human and robot demonstrations into interactive simulation gyms and brings skills acquired in simulation to physical robots. The Real2Sim module reconstructs editable scenes, aligns objects and cameras with the input, validates demonstrated or retargeted actions through native physics execution, and generates task-conditioned variations with action-feasibility checks. Within these environments, the agent generates executable code for manipulation stages, observes their outcomes, and distills successful attempts and failures into reusable task procedures, object-relative motions, and recovery strategies. Through a shared perception-and-control interface, these skills guide subsequent execution in simulation and on real robots, with motions adapted to current observations and no updates to the underlying model weights. Extensive evaluations demonstrate that Real2Gym enables high-fidelity simulation environment reconstruction, outperforming GPT-6 Astra Direct Mode by 16.7% in success rate with approximately 74.9% fewer policy-execution tokens across these environments, while exceeding it by 33.3% in physical robot execution success rate across four tasks on a real Franka robot.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37089
PDF
-
Single-turn emergency psychiatric triage across 15 frontier AI chatbots
2026-09-29 08:58:52Veith Weilnhammer, Lennart Luettgau, Christopher Summerfield, Raymond Dolan, Elise Wilkinson, Virginia Corno, Viknesh Sounderajah, Matthew M NourAbstract
People increasingly turn to general-purpose AI chatbots for advice about emotional and mental health problems, but the ability of these systems to recognize and appropriately triage psychiatric emergencies remains under-characterized. We evaluated psychiatric triage performance in 15 frontier AI chatbots using 112 clinical vignettes spanning four urgency levels, from routine care to immediate emergency assessment. In each trial (1680 total), a chatbot received a single user message conveying all triage-relevant information from one vignette and recommended a timeframe for care. The primary outcome was emergency under-triage; secondary outcomes included triage accuracy and the direction of errors. Vignettes and user messages were generated using a clinician-verified LLM pipeline. Across 415 emergency trials, 23 were under-triaged (5.5%; 95% CI 1.8-15.9). Overall accuracy, averaged across urgency levels, ranged from 42.0% to 71.8% across chatbots and was lowest for intermediate cases (19.6%; 95% CI 11.7-28.1). Every chatbot showed a net over-triage bias; overall, 763 of 786 incorrect assignments (97.1%) were more urgent than the prespecified triage level. The error pattern was similar when predictions were assessed against clinician ratings: 35 of 430 trials involving vignettes rated as emergencies by at least 75% of clinicians were under-triaged (8.1%). AI chatbots recognized most psychiatric emergencies but still missed clinically important cases and frequently over-triaged less urgent presentations. Further evaluations should examine how triage performance changes when clinically relevant information must be elicited through conversation.
Abstract (translated)
URL
https://arxiv.org/abs/2604.25415
PDF
-
Selecting The Most Informative Tokens in Natural Language Autoencoders
2026-09-29 08:50:02Federico Torrielli, Gianluca Barmina, Andrea Blasi N\'u\~nez, Amon Rapp, Luigi Di Caro, Peter Schneider-Kamp, Lukas Galke PoechAbstract
Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly. Which positions should an auditor inspect to understand a potential threat? We study this question across $4.7$ million explanations on prompt injection and concealment. We compare signals from model computation with a ranker trained only on chat structure. Chat structure usually selects more relevant explanations than the computational signals, without requiring a model forward pass for position selection. On three of four datasets, explaining just $5\%$ of positions retains nearly all of the success rate from explaining every position, where success means obtaining an explanation about the threat. The benefit varies with the audit task. We also show that pretrained verbalizers recover words that models have learned to conceal through fine-tuning, without additional verbalizer training. These results identify where auditors can concentrate explanation generation and show that useful explanations can extend beyond the model a verbalizer was trained to describe.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37040
PDF
-
MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation
2026-09-29 07:25:29Ke Wang, Houxing Ren, Zimu Lu, Yunqiao Yang, Zhuofan Zong, Mingjie Zhan, Hongsheng LiAbstract
End-to-end full-duplex speech models have brought open-source machine conversation closer to human-like interaction, yet existing systems remain limited in two intertwined dimensions: long-context robustness and multi-party interaction. Real-world scenarios such as meetings, group lessons, and social-robot reception require a single model to track, contextualize, and respond to multiple speakers over extended durations. Progress is constrained by both data and evaluation: open multi-party speech corpora remain small and are not designed for codec-frame-level full-duplex modeling, while existing long-audio benchmarks focus on passive listening and speech-to-speech benchmarks are mostly short and dyadic. We extend the Moshi paradigm jointly along the long-horizon and multi-party axes in English and Chinese. First, we release 57.6k hours of synthetic training data ($\href{this https URL}{MultiTalkPT}$ and $\href{this https URL}{MultiTalkFT}$) for long-form, multi-party, English-Chinese full-duplex dialogue, with controllable length, participant count, turn-taking, overlap, backchannels, interruptions, addressee shifts, and long-range coreference. Second, we introduce $\href{this https URL}{MultiTalkBench}$, built from real human recordings, for evaluating long-form, multi-party, bilingual full-duplex dialogue. Conversations average 32.6 minutes and include probes for long-range entity tracking, topic coherence, and addressee selection. Third, we train a bilingual Moshi-style model that sustains coherent multi-party English-Chinese conversations over extended durations and substantially outperforms open-source baselines including Moshi, MiniCPM-o-4.5, and Qwen3-Omni-30B-A3B-Instruct on MultiTalkBench.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36903
PDF
-
RareDx: Controlled Knowledge Integration and Graph-Grounded Policy Optimization for Rare-Disease Diagnosis
2026-09-29 07:14:46Bo Zhang, Yuchen Wang, Dongbai Li, Matthew Yu Heng Wong, Qingkai Zeng, Lijun Wang, Tien-Yin Wong, Peng Cui, Tianyu LiuAbstract
Rare-disease diagnosis is a long-tail reasoning problem: phenotypes are incomplete, individual disorders are sparsely documented, and relevant evidence is distributed across ontologies, gene annotations, and biomedical text. Language models consequently favor common conditions, miss rare candidates, or produce plausible but invalid names. We introduce RareDx, which couples controlled evidence use with knowledge-graph-grounded policy optimization. RareDx-Harness normalizes heterogeneous records into one ranked-diagnosis task and compares direct inference, static retrieval, adaptive tools, and structured phenotype-gene-disease reasoning over a shared knowledge layer. The training pipeline combines Top-10 post-training with RareDx-KGPO, our knowledge-graph-grounded policy optimization method. Its reward projects predictions into a canonical disease graph and integrates curated graded relevance, ontology proximity, biomedical similarity, and phenotype consistency. Vocabulary and output-budget constraints prevent dense partial credit from rewarding fabricated or overlong differentials. Across eight benchmarks, the complete RareDx system centered on Qwen3.5-9B reaches 38.34 macro Hit@10, 1.60 points above GPT-5.5 under the archived protocol; a disjoint validation-selection audit retains a 6.80-point routing gain over Direct on held-out cases. The 27B system reaches 23.53/36.56/40.76 at Hit@1/5/10. Controlled ablations show that retrieval is not uniformly helpful and that controlled routing is central to the gain. These results indicate that structured medical knowledge can turn a compact model into a competitive diagnostic ranker across heterogeneous long-tail settings in clinical practice.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35549
PDF
-
Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue
2026-09-29 06:59:37Omar Sheta, Rinku Deuja, Hadi Masoudi, Minghong FangAbstract
Safety-aligned language models are commonly deployed as multi-turn assistants, which lets adversaries spread unsafe intent across several user turns instead of a single prompt. Gradient-based jailbreak detectors such as GradSafe were developed for single prompts: they score an input by the alignment between its induced gradient and a fixed unsafe reference direction, and their effectiveness in multi-turn dialogue remains unclear. We conduct a controlled evaluation of gradient-based jailbreak detection in multi-turn settings. We extend GradSafe with a Context Window Scanner that applies the detector to fixed-size windows of user turns and uses the maximum window score as the conversation-level score. We evaluate different window sizes, attack families, benign conversation distributions, and target models. The results differ sharply between synthetic and realistic benign settings. Against synthetic benign conversations, the detector achieves an ROC-AUC of 0.98 on human-authored multi-turn jailbreaks. On WildChat benign conversations, ROC-AUC drops to 0.76, and a threshold calibrated on synthetic data flags more than 90% of benign conversations as unsafe. Under realistic benign distributions, single-turn windows give the highest separability, whereas longer windows and accumulated contexts reduce performance. The detector is also sensitive to the attack-generation method and target model: successful Crescendo attacks receive scores comparable to or lower than benign conversations, and Qwen2.5-7B-Instruct yields near-random separability with a different optimal window size. These findings show that gradient-based signals can support multi-turn jailbreak detection, but reliable deployment requires calibration on realistic benign conversations, short-window scoring, length-aware thresholds, and evaluation across attack types and model architectures.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36849
PDF
-
Adaptive Consistency Graph for Long-Horizon Agents
2026-09-29 06:28:56Jiecong Wang, Hao Peng, Zhanyi WangAbstract
Large language model agents can often make reasonable local decisions on short tasks, yet their performance degrades when success requires long sequences of dependent actions and tool calls. During execution, task requirements, historical evidence, and the current execution state may gradually become disconnected, so later decisions can drift from the original objective. We study this problem by introducing the Adaptive Consistency Graph (ACG) for long-horizon execution. ACG incrementally organizes execution evidence and its provenance in a persistent graph, then constructs a temporary requirement-centered view for each decision under a bounded context budget. Rather than replacing the base agent's planner or tool executor, ACG provides a structured and traceable context view for each decision. In the matched evaluation, ACG improves GPT-5.6-luna's average success from 44.5\% with ReAct to 50.2\%, with the largest gain on BrowseComp-Plus (73.5\% versus 62.4\%). We further analyze trajectory structure and inference cost to characterize this improvement. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32754
PDF
-
Spotter: Let the Embodied Model Lead, and the VLM Reflect for It
2026-09-29 06:21:50Long Li, Qichao Zhao, Yue Yang, Fan Xu, Zhe Wang, Alan Wee-Chung Liew, Chao Qu, Heng Tao Shen, Shirui PanAbstract
Current embodied models do not respond to their own failures, although what just went wrong could inform a small adjustment on the next attempt, the kind of reflection behind the gains of thinking in language models. We test whether they can repair a known error, which requires producing a correction and judging whether it is right. Stopped at a failure and allowed to retry, they seldom repair it through their own randomness or from a language description of the error, and best-of-N selection cannot pick the successful candidate after a failure. We attribute this to training only on successful demonstrations and to inputs too narrow to show what went wrong, and conclude that reflection must come from a vision-language model (VLM), which takes in far more information, such as the episode history and text, and is more general. Prior VLM-led work has the VLM plan every step and invoke the embodied model as a tool, placing the VLM on the critical path. We propose Spotter, which reverses the roles: the embodied model leads and executes continuously, while the VLM runs in parallel, monitors through a lightweight local screener, intervenes only when an error is detected, reflects on and corrects it, and returns control. We run Spotter with Qwen and with GPT as the VLM, and both improve the embodied models; with GPT, Spotter improves Cosmos Policy and $\pi_{0.5}$ by 5.6 and 7.5 percentage points on RoboCasa, and raises $\pi_{0.5}$ from 47.2% to 57.0% on the Hard setting of RoboTwin 2.0 and from 53% to 83% on a real robot. Because the VLM steps in only when an error is confirmed, a successful episode with Qwen takes only 13 to 16 s longer than with the embodied model alone and about 70% less time than with a VLM-led baseline using the same model. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36808
PDF
