MT‑WAM enhances the Fast‑WAM framework by adding complementary supervision for future 2‑D point trajectories and visual features while keeping the original training objectives. A lightweight dual‑stream branch and structured attention mask isolate motion‑specific processing, and motion‑stream tokens provide additional dynamics cues to the action expert. During inference, MT‑WAM skips future‑video prediction, using cached video and motion information to achieve higher success rates on LIBERO, LIBERO‑Plus, RoboTwin 2.0 Clean2Rand, and several real‑world tasks.
By Yiguang Yang, Jiankun Peng, Xiaoming Wang, Yiran Zhang, Zhibo Fang
The paper introduces a Chinese competitive debating dataset and benchmark, comprising 148 fully adjudicated matches with 2,698 stages and 20,542 exchange units, all annotated by 120 professional judges using a shared rubric. It offers manually verified transcripts, segmentation, and preserved scores, votes, and rationales, enabling evaluation of large language models on winner-tendency prediction, stage-score prediction, and best-debater prediction. Zero‑shot experiments show the best models achieve 66.2% winner‑prediction accuracy, a 0.250 Pearson correlation for stage scores, and 56.8% best‑debater accuracy.
By Zongrui Yang, Haoyuan Li, Zhongsheng Wang, Zhirui Zeng, Pengqian Han, Yi Zhou, Yuting Wang, Jiamou Liu
Samsone is a family of small audio language models (SALMs) designed for on‑device inference, with the flagship Samsone‑134M setting a new state‑of‑the‑art for its size class across multiple benchmarks. The paper also presents Samsone‑99M and Samsone‑356M to study scaling laws, showing that these compact models achieve performance competitive with much larger counterparts. The authors train the models on publicly available data and release training code, weights, mobile‑optimized checkpoints, and an open‑source Android app for real‑time inference.
By Piotr Masztalski, Micha{\l} K. Grzeszczyk, Olaf Sikorski
CIPL (Channel Inversion for Privacy Leakage) is a channel-aware framework designed to evaluate black-box privacy leakage in large language model agents. It models the leakage process through stages of sensitive source, selection, assembly, execution, observation, and extraction, assessing how selected sensitive units become attacker-recoverable outputs. Experiments across memory, retrieval, and tool-mediated targets, plus a live-agent case study, reveal that recoverability depends on factors beyond storage labels, such as observation surface, prompt alignment, retrieval depth, and provider behavior, and that a semantic audit can uncover disclosures missed by exact matching.
By Tao Huang, Guosen Wu, Guolong Zheng, Jiayang Meng, Chen Hou, Xu Yang, Xuechao Yang, Feng Xia
The paper introduces a new multi-draft speculative sampling algorithm that uses Poisson processes to improve inference efficiency and output provenance for large language models. It achieves strong sampling efficiency while allowing an unbiased watermark to be embedded without reducing speculative acceptance. The method relies on an exact list‑coupling‑without‑communication scheme, giving a drafter‑invariant property that benefits both sampling and watermarking, and the authors experimentally confirm its effectiveness.
By Yanxiao Liu, Sicheng Wan, Zhan Gao, Deniz G\"und\"uz
ProTracer is a training‑free framework that uses Vision‑Language Models (VLMs) together with proprioceptive signals to analyze robot manipulation failures. It performs binary failure detection, categorization, explanation generation, and introduces failure onset localization—identifying the earliest moment a robot deviates from a valid trajectory leading to failure. The method leverages proprioceptive dynamics to pinpoint informative action boundaries and converts robot‑state signals into natural‑language descriptions for joint multimodal reasoning, achieving strong performance on both conventional failure diagnosis and the new failure onset localization task.
By Chang Dong, Mehdi Hosseinzadeh, King Hang Wong, Lingqiao Liu, Francois Fraysse, Feras Dayoub, Minh Hoai Nguyen
The paper investigates whether machine learning models can uncover physical patterns in atomistic data without relying on traditional physics-based inductive biases such as geometric locality or graph structures. By training a general-purpose architecture on molecular simulation data, the authors demonstrate that the model autonomously learns interatomic interaction strengths resembling classical electrostatics and identifies interaction cutoffs aligned with established physical models. The study also reports predictable neural scaling behavior and competitive accuracy on certain metrics compared to physics-informed architectures, suggesting that explicit priors may only be necessary when empirically justified.
By Tobias Kreiman, Yutong Bai, Fadi Atieh, Elizabeth Weaver, Eric Qu, Aditi S. Krishnapriyan
Fact Grounded Attention (FGA) is a new architectural change that injects verifiable knowledge directly into the transformer’s pre‑softmax attention scores, turning large language models into deterministic truth‑tellers. By modifying the attention mechanism itself, FGA prevents hallucinations for facts that exist in its knowledge base, unlike prior methods that patch output or prepend retrieved text. Experiments on 1,107 technical queries show accuracy rising from 6.3% with vanilla Llama 3.2 to 99.7% with FGA, and knowledge updates can be applied in under one second without retraining.
By Aayush Gupta, Manish Choudhary
The paper introduces a training strategy for cascaded simultaneous speech translation that allows the system to dynamically decide how much of the source prefix to translate. By fine‑tuning a large language model (Qwen3‑8B) on stable prefixes—pairs of source prefixes and the longest shared translation with the full sentence—the authors enable contextual read‑write decisions beyond fixed wait‑k or target‑suffix deletion. Experiments on English‑to‑German, Japanese, and Chinese demonstrate that stable prefixes improve the quality‑latency tradeoff across various test sets.
By Hieu Hoang, Amittai Axelrod, Matt Post
The paper introduces Bayesian Chronicle Agents (BCA), a lightweight belief layer that separates an LLM agent’s internal stance from its outward speech. Each stance is represented as a probability updated via a single Bayesian step per utterance, with a single prior‑strength parameter κ controlling stubbornness. By sweeping κ, the authors generate three controllable opinion‑dynamics regimes—consensus, persistent disagreement, and committed‑minority influence—matching Friedkin–Johnsen theory and demonstrating recoverable, auditable belief states across models.
By Hafsa Akbar, Daniel Platnick, Marjan Alirezaie, Hossein Rahnama
The study evaluates how large language models (LLMs) respond to the Norwegian Moral Foundations Questionnaire (MFQ‑30) and compares their moral profiles to a Norwegian human sample. Six open‑weight LLMs were tested, with half engaging with the questionnaire and the other half producing flat, near‑human responses. Two steering methods—prompt‑level persona steering and activation‑level ActAdd—were applied; a neutral Nordic persona improved alignment with the Norwegian mean by 44‑77%, while ActAdd flattened profiles without targeting specific foundations.
By Hans Andersen, David Dichas
The paper introduces a three‑stage pipeline to improve accented conversational ASR for speakers from India, Indonesia, and Latin America. It uses heuristic SQL filters to curate entity‑rich training data, regional LoRA adapters fine‑tuned on Qwen2.5‑Omni‑3B to generate both verbatim and corrected transcripts, and a six‑category error taxonomy validated by an LLM judge. The approach raises entity recall to 80‑85% and filler recall to 76‑86%, while keeping WER low (6‑10%) and outperforming Whisper and a commercial ASR on entity recall.
By Fiza Husain, Ankit Pandey, Yash Singh
VISPATH is a visual‑intent‑guided path reasoning framework designed for multimodal knowledge graph question answering (MM‑KGQA). It first identifies a reliable starting entity by fusing multimodal grounding with graph‑structural cues, then iteratively discovers and refines reasoning paths using hop‑specific multimodal intent and a reasoning‑chain pruning step. The framework is evaluated on the newly introduced VISPATH‑Bench, which tests two‑to‑four‑hop reasoning, and demonstrates consistent improvements over strong baselines, even surpassing GPT‑5.4 when using GPT‑4o as the backbone.
By Jinke Wu, Zhengpin Li, Mengzhe Jia, Yang Li, Wentao Zhang
PhysioBench is a unified benchmark for physiological signal question answering that consolidates annotations from 22 public datasets into 61.4 million question‑answer pairs spanning 30 tasks. Each pair is linked to a specific signal segment and traceable to its source annotation. The benchmark evaluates 21 models—including large language models, vision‑language models, time‑series language models, and physiological signal foundation models—under three settings, revealing that no model consistently excels across all modalities and tasks, and that performance is sensitive to question wording.
By Mengxuan Li, Junfa Chen, Jinze Xia, Yundan Chen, Lixin Fan, Ke Liu, Keyue Shi, Haishuai Wang
ServeGuard is a supply‑chain primitive that allows a publisher to ship a proof‑carrying adapter for an open‑weight language model, proving in zero‑knowledge that the adapter contains no hidden backdoor channel in the monitor’s blind subspace. The proof is inexpensive because it relies on a deterministic function of the public base model, and the served residual is the model’s own public floor. The system lets consumers or regulators verify the absence of this class of hidden channels without revealing the certified read factor or trusting the publisher.
By Dominik Dahlem, Rui Vieira
The study evaluates six advanced language models on a two‑agent <log_2 N>‑Questions game, where a questioner must identify a secret Wikipedia paragraph using exactly <log_2 N> binary questions answered by an agent that only sees the target. Across 408 games, win rates decline geometrically with horizon length (p≈0.93), and per‑round failure rates remain flat, indicating that errors compound because more rounds must succeed rather than because individual rounds become harder. Adjudication reveals that losses stem from single‑agent answer errors and discrimination failures, with Claude Opus 5 lagging due to high false‑negative rates, while the top five models cluster closely; maximizing information gain requires structural partitioning, and neither reasoning‑token usage nor API cost correlates with success.
By Peter Potash
RheoSampling tackles the one‑hot dilemma in stochastic dynamic‑tree speculative decoding by decoupling token sampling from tree construction and verification. It injects a sampled token into deterministic top‑K slots, assigning distinct proxy probabilities for expansion and pruning while preserving the true sampling probability for verification, thereby achieving lossless, context‑aware top‑K construction with stochastic sampling. Experiments on various LLMs show higher acceptance rates and speedups compared to existing dynamic‑tree methods.
By Qiao Hu, Yepeng Weng, Bo Zhang, Takehisa Yairi
HERMES is a holistic end‑to‑end multimodal driving framework that incorporates long‑tail semantic knowledge into trajectory planning for autonomous vehicles. It uses a foundation‑model‑assisted annotation pipeline to build Long‑Tail Scene Context and Long‑Tail Planning Context, capturing hazard‑centric scene information, maneuver intent, and risk‑aware guidance. A Tri‑Modal Driving Module then fuses multi‑view visual observations, historical ego‑motion, and long‑tail semantic instructions to generate intent‑ and risk‑aware trajectories, achieving consistent performance gains on a large‑scale real‑world long‑tail driving benchmark.
By Weizhe Tang, Junwei You, Jiaxi Liu, Zhaoyi Wang, Rui Gan, Zilin Huang, Feng Wei, Bin Ran
The paper presents a three‑stage, parameter‑efficient approach to improve automatic target recognition (ATR) with synthetic aperture sonar (SAS) data by adapting DINOv3 Vision Transformers. Stage 1 applies Low‑Rank Adaptation (LoRA) while freezing the backbone, which significantly boosts the area under the precision‑recall curve from 0.300 to 0.679. Subsequent hard‑negative mining and supervised contrastive learning stages show negligible impact, indicating that a single LoRA adaptation is sufficient for effective underwater ATR.
By Dan Zimmerman, Frank E. Bobe III, Amelia L. McCormack, Matthew Cook, Gregory D. Vetaw
arXiv:2609.21672v1 Announce Type: new
Abstract: Large language models (LLMs) achieve strong performance but suffer from slow and costly inference. Existing acceleration methods often lead to noticeab...
By Zhenyu Zhang, Jiudong Yang, Zhaowen Tao, Meng Chen