The paper introduces Adaptive Self-Consistency (ASC), a method that treats large language models as grey-boxes by leveraging the full answer distribution from log‑probabilities rather than single sampled answers. It formulates inference as sequential mode identification with distribution‑valued observations, proving that this approach never performs worse than traditional black‑box sampling and can stop earlier. The proposed ASC‑D algorithm, a betting‑based stopping rule, achieves significant reductions in required trajectories—up to 95.6% fewer—while attaining the best fixed‑budget correct‑certification rates on MMLU‑Redux across three open‑source models.
By Jingkai Huang, Yunfan Zhang, Will Ma, Weihua Zhou, Zhengyuan Zhou
DrivingBench is the first benchmark that tests general‑purpose vision‑language models on the task of driving a real Toyota Corolla around a parking‑lot cone course. The models receive live camera frames and issue steering and velocity commands, with inference latency counted as part of the challenge. In tests, only GPT‑6 Astra completed the course, while other models showed limited progress or failed to pass half the course.
By Aditya Ramabadran, Simon Mahns, Tobias Gessler
APTInvestBench is a benchmark that evaluates how well autonomous agents can investigate advanced persistent threats (APTs) when faced with different telemetry settings. It contains 370 cases derived from 56 attack reconstructions, totaling 16.4 million log records, and tests agents on seven SOC-inspired telemetry conditions. The benchmark measures evidence acquisition and formal citation support, revealing that while overall coverage drops only slightly when telemetry is limited, a significant portion of actions lose sufficient citation support, highlighting instability in agent performance.
By Yu Wang, Shuhao Li, Tao Yin, Ziyang Li, Xueying Zhao, Peishuai Sun, Jiang Xie
The paper introduces TED (Text-Axis Evidence Decomposition), a post‑hoc scoring method that improves anomaly localization in CLIP‑based detectors without altering the backbone or prompts. TED evaluates whether ambiguous responses are better supported by defect patches or normal patches, thereby distinguishing true defects from visually complex normal regions. Experiments show that TED significantly enhances pixel‑level localization across frozen VLM backbones and adapted hosts, especially under hard‑false‑positive competition.
By JinYoung Kim, Geonho Kim, GiJeong Park, Geonu Lee, YoungJoon Yoo
The paper introduces an LLM-based framework for continuous dimensional emotion evaluation in multimodal dialogue, combining discrete emotion recognition with Valence-Arousal-Dominance (VAD) assessment on the IEMOCAP dataset. It incorporates acoustic cues as natural language descriptions via the SpeechCueLLM approach and evaluates six models from the LLaMA, GPT, and Qwen families using zero-shot, few-shot, and LoRA fine-tuning. LoRA-fine-tuned LLaMA models outperform prompt-engineered GPT models, achieving a new state-of-the-art Valence CCC of 0.7822, and ablation studies show that textual audio descriptions significantly benefit smaller models.
"whyItMatters":"The study demonstrates that domain adaptation through fine-tuning can surpass larger GPT models in multimodal emotion evaluation, highlighting the importance of tailored training for emotion recognition tasks."
By Yutong Hu, Jinho Choi
The paper introduces HealBench, a benchmark of 265 runtime errors from 18 real-world repositories, and HealGuard, a safety framework that restricts healing code to an analyzable subset of Python and applies static and dynamic taint analysis. Using these tools, the authors evaluate a dedicated healing method and three general coding agents powered by different LLM backbones, achieving a 38.11% resume rate and a 28.68% test‑pass rate, while HealGuard flags 17.4% of successful healings as potentially unsafe. The study demonstrates that current LLM agents can meaningfully repair real repository crashes, but also highlights significant safety concerns that the Guardrail framework can detect, albeit with a high false‑positive rate.
By Gou Tan, Pengfei Chen, Zhensu Sun, Jieke Shi, Junkai Chen, Ting Zhang, Weifeng Sun, Junda He, Shuai Liang, Chuanfu Zhang, Lwin Khin Shar, David Lo
DAGent introduces an Evaluate‑then‑Grow planning approach for deep research agents, building directed acyclic graphs incrementally based on confidence and uncertainty from completed tasks. The framework includes a hierarchical context layer for efficient query handling and a structural reinforcement learning component, DAGRPO, that rewards topology‑conditioned execution. Experiments on BrowseComp‑Plus, GAIA, and xbench‑DeepSearch show DAGent outperforming strong baselines across multiple backbones and scaling to large language models.
By Hanwen Liu, Yuanfu Sun, Qiaoyu Tan
UniAE-MoE is a unified audio encoder that uses a Mixture-of-Experts architecture to model cross‑domain audio representations. It integrates encoder components from Qwen2‑Audio and Audio‑Flamingo 3, enhances them with SwiGLU and shared experts, and applies a two‑stage instruction‑tuning strategy along with task‑specific data scaling. The model achieves state‑of‑the‑art results on the XARES‑LLM benchmark (0.802) and tops the Interspeech 2026 Audio Encoder Capability Challenge, demonstrating strong generalization across speech, music, and general audio tasks.
By Shengbo Cai, Zhisheng Zhang, Zichao Nie, Jing Peng, Jingran Xie, Zhiyong Wu
EngramBench is a new benchmark designed to evaluate skill evolution in autonomous agents by focusing on genuine capability abstraction rather than solution copying. It includes 30 learning tasks and 13 unseen transfer tasks that require agents to manage complex, multi-hour development cycles with LLM‑simulated users. The study shows that while static skill banks cannot eliminate the need for precise code implementation, they effectively reduce redundant context and cut overall coding time by more than 55%.
By Zhixuan Tan, Pengjie Gu, Zhao Li, Yihan Hu, Xu He, Dong Li, Jianye Hao
The paper introduces FDCU, a dual‑constrained subspace projection framework designed to improve machine unlearning for large language models. FDCU limits parameter updates with a dual‑masking rule that preserves general knowledge via Fisher Information while preventing the activation of spurious suppressors through the Principle of Minimal Functional Intervention. Experiments show that FDCU achieves state‑of‑the‑art robustness against retraining attacks while maintaining near‑lossless general utility, thereby ensuring durable safety for LLMs.
By Jiaqing Li, Shide Zhou, Zhibo Zhang, Yuxi Li, Tianlong Yu, Kailong Wang
The paper introduces Galahad, a memory layer that stores a transformer language model’s key‑value state for blocks of text, allowing subsequent requests to reuse previously computed attention rather than recomputing it. On seven real‑world datasets, 98.7% of prompt tokens were already read, and with Galahad the model could attend to an entire 97,000‑token corpus, achieving 98–100% recall on a 100‑fact test while reducing inference time and energy consumption dramatically. The approach was validated across 30 models and all runtimes, demonstrating that stateful inference can replace stateless serving without loss of accuracy.
By Sietse Schelpe
The paper introduces a new skill poisoning technique for large language model agents that decouples the pretext (rationale) from the actuation (operation). By separating these two risk‑realization factors, the authors create coordinated pretext‑actuation skill pairs that allow malicious actions to remain hidden within legitimate agent behavior. An automated framework is presented to discover execution dependencies, synthesize these skill pairs, and refine them through closed‑loop feedback, achieving high attack success in both single‑session and persistent scenarios.
By Wenxin Wu, Lingyong Yan, Lei Sha, Shuaiqiang Wang, Jiashu Zhao
The paper introduces STAR‑Ar, a BERT‑BiLSTM‑CRF model designed for the Daleel 2026 Arabic argument mining shared task. It treats argument discourse unit detection and classification as a token‑level sequence labeling problem, achieving an F1‑score of 72.69 on validation and 73.7 on test data. Analysis shows that models trained only on editorial texts perform worse than those trained on debates, mainly due to the smaller editorial dataset.
By Bhuvanesh Verma, Ali Abusaleh, Alexander Mehler
GroundingPI is a 4‑billion‑parameter grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. It is trained with multimodal and spatial pretraining, supervised fine‑tuning, and reinforcement learning, achieving a new state‑of‑the‑art average of 73.68% across 34 grounding benchmarks. As a visual backbone, GroundingPI improves performance in robotic manipulation and autonomous driving, outperforming larger models and mainstream backbones in several out‑of‑distribution settings.
By Qize Yu, Lianrui Fan, Boyu Chen, Jiaqi Liang, Xini Ding, Yue Chen, Zetian Song, Yuran Wang, Yi Zou, Kaixuan Wang, Tianxing Chen, Wenxuan Song, Bohan Zhou, Mingleyang Li, Siqiao Huang, Yuqi Ye, Caigao Jiang, Wei Wei, Ruihai Wu, Hang Zhang, Yixiao Ge, Shuchang Zhou, Shilong Liu, Xianming Liu, Ping Luo, Shiyu Huang
D‑Scope is a framework that links the interpretation of sparse autoencoder (SAE) features in diffusion transformers (DiTs) to controllable image generation. It aggregates SigLIP‑2 embeddings of highly activating image patches into visual centroids, matches target text descriptions against these centroids, and retrieves individual features without per‑feature text annotations. The method provides visual evidence for each selection and uses spatially masked interventions to test decoder directions under fixed generation conditions, evaluated across 150 SAEs and a benchmark of 100 target concepts.
By Xinyue Xu, Jiahao Zhang, Lijie Hu, Peter Hase, Hao Wang
The paper titled "Pretext: Defeating Malicious Skill Detection Frameworks for AI Agents" demonstrates how an attacker can bypass current skill‑scanning defenses by crafting malicious skills that evade both static analysis and LLM‑based semantic checks. By moving malicious payloads into natural language and distributing instructions across files, the white‑box attacker named Pretext achieves high evasion rates—up to 97% against a frozen detector and 77% against a co‑adaptive one—across three open‑source models.
By Tobias Kaisar, Aritra Dhar
The paper examines the reliability of lie detection probes for language models when the models adopt anti-factual personas, such as conspiracy theorists. A dataset of 8,916 human-reviewed responses from three LLMs was created, and eight existing probes were evaluated, revealing many fail to flag falsehoods under these personas. The authors also constructed confounder datasets showing that probes often track spurious correlations like instruction compliance, and propose a simple linear probe that performs best on both persona and confounder tests.
By Maximilian von Klinski, Sebastian Lapuschkin, Wojciech Samek, Lennart B\"urger
Mid‑Harness proposes a test‑time compute strategy that samples and verifies candidate actions before execution, keeping the underlying generator and harness unchanged. Experiments show that with a strong verifier, sampling more actions significantly boosts success rates—e.g., a GPT‑5.6 verifier raises Pass@1 from 50.00 % to 68.03 % on TerminalBench‑Lite using eight samples. The approach also improves performance across various models, benchmarks, and harnesses, demonstrating that action scaling is a promising target for enhancing terminal agent reliability.
By Minki Kang, Ryo Hachiuma, Shaokun Zhang, Subhashree Radhakrishnan, Yonggan Fu, Jindong Jiang, Mingjie Liu, Ehsan Hosseini-Asl, Yi Dong, Yu-Chiang Frank Wang, Byung-Kwan Lee
PhantomEnvironments is a framework that trains large language model agents in synthetic, rule‑generated fictional worlds. By creating multi‑turn reinforcement learning environments where agents search templated articles to answer multi‑hop questions, the approach eliminates the need for costly human data or hallucinated LLM‑generated settings. Agents trained in these zero‑cost, purely rule‑based worlds transfer effectively to real‑world multi‑hop search benchmarks, often surpassing models trained on real data, and demonstrate scalable search behavior that grows linearly with question difficulty.
By Anmol Kabra, Swathi Saravana Selvam, Albert Gong, Chao Wan, Christian Belardi, Dongyoung Go, Katie Z. Luo, Kilian Q. Weinberger
arXiv:2609.40322v1 Announce Type: cross
Abstract: Material generation should produce not only an appearance, but also the rules that construct it. We introduce MatLoom, a compact, layer-oriented lang...
By Anson Y. Lam, Shuqing Li, Michael R. Lyu