The paper introduces TRACE, a token‑level objective designed to reduce multi‑turn safety risks in large language models. TRACE assigns each token a weight based on the discounted return of a refusal‑attributable advantage, comparing a frozen reference model with a refusal‑ablated copy to credit early tokens for later refusal evidence. Evaluated across five open‑weight models and seven multi‑turn attacks, TRACE achieves the lowest attack success rate in all 35 model‑attack pairs while maintaining model utility within 1.23 points on MMLU and HellaSwag.
By Fengpeng Li, Kemou Li, Qizhou Wang, Haiwei Wu, Jiantao Zhou, Di Wang
The paper introduces Spiking Contrastive Attention (SCA), a module designed to reduce spectral bias in Spiking Transformers by enhancing high‑frequency information. It demonstrates that spiking neurons and spiking self‑attention act as low‑pass filters, leading to loss of high‑frequency components. Experiments show that SCA improves performance across image classification, semantic segmentation, and event‑based tracking while maintaining lower complexity than the original spiking self‑attention.
By Xiaoli Liu, Malu Zhang, Yang Yang
DRelay introduces a global draft context mechanism to improve prefix-aware parallel speculative decoding for large language models. By using a global reader to extract predictive information across the entire draft block and a causal selector to repair early token selection errors, DRelay extends the accepted prefix length and enhances decoding performance. Experiments on eight benchmarks show consistent gains over existing methods such as DFlash, Domino, and DSpark, with notable speedup improvements in SGLang serving.
By Zhuoyu Wang, Junnan Huang, Xinyu Chen
The paper introduces a multi‑agent large language model framework that interprets personalized health checkup results by assigning distinct tasks to specialized agents and combining their outputs. Experiments on 120 Korean compound queries show that the multi‑agent approach improves overall quality scores and user preference compared to a single‑agent baseline, though it incurs higher latency and cost. The benefits are most pronounced for queries requiring personal‑record lookup.
By HyungJun Kim, Taehan Lee, Soojin Cheon
The paper audits whether routing entropy in Attention‑Residual transformer variants (Swin‑Tiny and DeiT‑Small) trained on CIFAR‑10/100 can signal prediction uncertainty beyond model confidence. Three tests examine the presence, consistency, and predictive power of routing signals, while a sensitivity audit measures how much injected effect the probes recover. Results show no significant improvement over confidence alone, with only modest recovery of injected signals and no consistent gains across seeds or metrics.
By Wenhao Liang, Lin Yue, Wei Emma Zhang, Mingyu Guo, Olaf Maennel, Weitong Chen
OpenMTB‑Audit is an open‑source benchmark that tests large language models on 500 synthetic non‑small cell lung cancer cases, covering five adversarial error categories and four safety labels: Supported, Partially Supported, Unsupported, and Insufficient Information. The study found that all eight tested LLMs over‑refused Partially Supported recommendations, collapsing labels to achieve high safety scores. A deterministic seven‑module framework, MTB‑AuditAgent, was introduced to reduce over‑refusal to 6.7% and reach 91.2% accuracy, while an oncologist annotation study highlighted disagreement around the boundary between information sufficiency and treatment optimization.
By Negin Ashrafi, Jia Luo, Stacey M. Frumm, Roxana Daneshjou
The paper evaluates five vision‑language models on autonomous driving tasks under various visual input conditions, finding that visual corruption affects accuracy and confidence differently across models and datasets. It then tests Visual Evidence Augmentation (VEA) as an inference‑time technique to enhance reliability, observing mixed improvements depending on the model and setting.
By Manasa Mariam Mammen, Priyanka Mary Mammen, Zafer Kayatas, Stefan Wagner
The paper proposes a new framework for collaboration between a small language model (SLM) and a large language model (LLM) that treats the interaction as an information acquisition problem under an API budget constraint. Instead of delegating reasoning tasks, the SLM remains the primary reasoner and selectively queries the LLM advisor with targeted questions, using a three-stage RLVR approach to decide when to call the advisor, how to phrase queries, and how to integrate the responses. Experiments on mathematical reasoning and coding tasks show that this strategy improves the performance–cost tradeoff compared to existing baselines and can transfer to other advisor model families without additional training.
By Yongjun Kim, Xiaoxiao Li, Jaeho Lee
LineupRL introduces a reinforcement learning framework with verifiable rewards for time series captioning, using a frozen large language model to identify the correct time series from a set of distractors based on a generated caption. This approach bypasses the limitations of supervised fine‑tuning and traditional RL rewards that poorly transfer to open‑ended time series generation. Experiments on two captioning benchmarks, as well as forecasting and reconstruction tasks, show that LineupRL outperforms both SFT and RL baselines across all metrics, and its trained 3B vision‑language model surpasses a 72B model distilled from SFT captions. The method also demonstrates resistance to reward hacking and produces captions that accurately trace trends and name key values.
By Haochen Zhang, Laura Yao, Zachary Plotkin, Gengwei Zhang, Tianlong Chen
The paper introduces a continuous evaluation framework that assesses both outcome-level and process-level aspects of evolving enterprise AI agent skills. It applies this framework to two variants of a Business Value Determination skill, running 240 trials across multiple models, harnesses, and specifications. The results show that while most trials pass final numerical checks, a large majority still exhibit process-level deviations, and dependency attribution reduces the number of failed checks per run. The framework also provides reusable regression tests and highlights specification sensitivity across configurations.
By Ngoc Phuoc An Vo, Aarya Doshi, Vadim Sheinin
AVSD-Scenes is a new dataset of 12,291 audio‑visual scene descriptions for urban environments, built from the TAU Urban Audio‑Visual Scenes dataset. The descriptions are generated by first creating modality‑specific text with Qwen2‑Audio‑7B and Qwen2.5‑VL‑7B, then merging them with large language models (Qwen3‑14B, Mistral‑Small‑3.2‑24B‑Instruct‑2506, Gemma‑3‑27B‑it) to produce multimodal narratives that combine auditory and visual cues. Benchmarks show that these multimodal descriptions improve semantic alignment, cross‑modal retrieval, and scene classification accuracy (up to 95.4%) while remaining discriminative even without explicit scene labels.
By Dhanunjaya Varma Devalraju, Arshdeep Singh, Mark D. Plumbley
The paper surveys recent advances in Retrieval Augmented Generation (RAG), a technique that integrates external retrieval into language model generation to reduce hallucinations and keep knowledge current. It introduces a four‑axis taxonomy—efficiency, defense, interactivity, and reasoning—to organize contemporary RAG research, covering retrieval methods, fusion strategies, embedding optimizations, and reinforcement learning policies. The survey also reviews evaluation practices, domain‑specific applications, and architectural variants, while highlighting ongoing challenges such as retrieval quality, reliability, domain adaptation, scalability, and explainability.
By Meghana Sunil, Shravya V, Shravan Venkatraman, Joe Dhanith PR
The study evaluates counterfactual bias in ten open‑source large language models (LLMs) for pediatric Emergency Severity Index (ESI) prediction. By creating paired clinical vignettes that differ only in demographic or socioeconomic variables, the authors measure shifts in acuity assignment, finding that counterfactual sensitivity varies widely across model families and sizes. A fine‑tuned Qwen2.5‑7B model exhibited the lowest sensitivity, while larger or medical‑domain models sometimes showed greater shifts, highlighting the need for fairness assessment before clinical deployment.
By Manar Aljohani, Brandon Ho, Kenneth McKinley, Dennis Ren, Xuan Wang
The paper introduces Causal Memory Policy (CMP), a framework that identifies the utility of memories in memory‑augmented language models by intervening on retrieval rather than on storage. CMP reserves fixed context slots for memories sampled with known propensities and estimates utility using self‑normalized inverse propensity weighting, providing unbiased estimates and exact variance. Experiments show that CMP improves discrimination between required and non‑required memories and reveals that identified utility alone is insufficient for retention decisions across unseen queries.
By Arman Behnam, Binghui Wang
The paper introduces Bayesian Dialectical Argumentation (BDA), a method for aggregating answers from multiple large language models (LLMs) in a council setting. BDA treats each LLM’s typed moves—proposals, challenges, and concessions—as evidence in a classical annotator model, estimating per-agent reliability even when some agents are persistently unreliable. By weighting evidence according to these inferred reliabilities, BDA produces calibrated posterior probabilities for candidate answers and can invert unreliable agents instead of merely outvoting them, achieving superior calibration and robustness on both binary and multi-class benchmarks without extra LLM calls.
By Ionel Eduard Stan, Paolo Napoletano
SPHERE is an adaptive VR indoor scene generation framework that turns isolated 3D synthesis into continuous human‑AI co‑creation. It learns persistent spatial preferences from multimodal user interactions, abstracts these into hierarchical constraints for geometric resilience, and employs a human‑in‑the‑loop reinforcement learning loop to refine retrieval policies. A mixed‑design study with 42 participants and offline ablation show that SPHERE reduces corrective edits and physical effort while avoiding bias toward shallow object‑level traits, producing geometrically resilient, profile‑aligned layouts.
By Hyeonmin Lee, Zheng Wei, Kyungmin Kwon, Jumin Seo, Jiwon Park, Hayoung Oh
Mimir is a physics‑grounded large language model agent designed for long‑horizon irrigation control. It operates on two timescales: a fast scale that uses a structured physical interface and deterministic simulator to validate and refine LLM proposals before execution, and a slow scale that consolidates recurrent failure patterns into persistent contextual principles. Across multiple sites, crops, and years, Mimir achieves the lowest aggregate control cost and reduces irrigation usage by about 51% compared to historical schedules, while ablation studies confirm the importance of forward simulation, verified revision, and persistent context.
By Yimeng Liu, Mi Zhang, Younsuk Dong, Zhichao Cao
The paper proposes a hidden‑state probing method for detecting hallucinations at the span level in large language model outputs, moving beyond token‑wise binary classification. By examining layer‑wise activation patterns, the approach identifies the exact onset and continuation tokens of hallucinations, achieving higher precision‑recall AUC than random baselines despite class imbalance. Additionally, the authors introduce a cross‑model detection framework where one model observes another’s internal representations, showing that an external observer can match or surpass the generator’s own self‑detection of hallucination onsets, even when the observer is smaller.
By Kingshuk Gupta, Davide Buscaldi
The paper introduces VISTA, a visual harness that equips a general-purpose multimodal model with long‑horizon vision and a lossless visual memory. VISTA enables the model to directly perceive and actively retrieve past observations, allowing it to reorganize visual input during reasoning. On the ARC‑AGI‑3 benchmark, VISTA boosts Claude Opus 5.0’s Relative Human Action Efficiency from 40.68 to a perfect 100.00, completing all 25 public games with 57.4% fewer actions than first‑time human participants, and it also outperforms baselines on three additional visual game and puzzle benchmarks.
By Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu, Kaiming He
The paper demonstrates that reading a large language model (LLM) judge’s verdict from the logits of its first generated token—an approach used in constrained decoding and likelihood‑scoring evaluation—introduces a significant distortion in position bias. Because judges do not always start with a verdict token (12–49% of cases for Qwen3 judges and <3% for Llama‑3.1‑8B and Phi‑3.5‑mini), this readout often returns the first response rather than a true judgment, inflating position bias by up to 42 points while barely affecting judge accuracy. The authors recommend reporting the frequency with which a judge leads with a verdict token to provide a more accurate assessment of position bias.
By Gnaneswar Villuri, Hashmath Shaik, Alex Doboli