OmniHarness is a framework that enables generalizable visual generation by learning symbolic policies from verified executions. It abstracts shared procedures and applicability conditions, allowing these policies to be instantiated, adapted, and composed for new tasks while keeping model parameters fixed. The system uses intermediate verification for refinement, self-directed inquiry to generate practice tasks, and continuous feedback to expand capabilities, achieving strong results on multiple benchmarks and outperforming baselines on Creative tasks.
By Xu Xu (Beihang University), Jinxiu Liu (The Chinese University of Hong Kong), Zhangbo Qiao (Beihang University), Jiaxing Lu (Beihang University), Xiangyu Zhang (Beihang University), Yubin Gu (National University of Singapore), Fangwei Ning (Beihang University), Yan Shi (Beihang University)
MUUNRiver-Bench is a diagnostic benchmark for music retrieval that uses natural‑language instructions to define relevance for reference‑audio queries. It contains 3,440 tracks across 13 genres and 116 sub‑genres and covers seven tasks such as similar‑music, style‑preserving lyric‑rewriting, cover, and segment retrieval. Experiments with six models in eight configurations show that acoustic encoders favor local identity while text‑aligned encoders favor semantic relations, and that instruction‑aware and audio‑text fusion systems do not consistently outperform their backbones.
By Zhancheng Guo, Congren Dai, Shangda Wu, Jianhuai Hu, Danni Zhao, Xiaobing Li, Maosong Sun
RAG-CT is a defense mechanism designed to protect Retrieval-Augmented Generation (RAG) systems from leaking personally identifiable information (PII). It works by detecting malicious queries through analysis of entropy and margin distributions, applying a score-based detection method. Experiments across four attack strategies and four baseline defenses on two datasets show that RAG-CT significantly reduces PII leakage while outperforming existing defenses, all without altering the underlying LLM or retriever.
By Xingyu Lyu, Jiayimei Wang, Jianfeng He, Ning Wang, Yidan Hu, Yimin Chen
Auto-HSI is a system that creates personalized human‑swarm interaction interfaces on demand using large language models to automatically generate code from natural language descriptions and gesture demonstrations. The prototype enables untrained operators to control a swarm of 50 simulated robots with one‑ or two‑hand gestures, allowing teleoperation of motion, formation shape, and shape deformation. Experiments demonstrate the system’s gesture tracking, code generation, and live operation capabilities, including real‑time updates and deployment on real robots.
By Alessandro Nazzari, Nathan Cerisara, Dorian Tonnis, Raina Zakir, Lorenzo Labarile, Weixu Zhu, Marco Dorigo, Mary Katherine Heinrich
The paper introduces ASURRE, a benchmark dataset for detecting AI‑assisted responses in online surveys. It evaluates how different LLM usage strategies—ranging from full generation to persona‑grounded agentic completion—affect the performance of existing machine‑generated text detectors. The study finds that while naive AI usage is easily detected, more sophisticated persona‑grounded agents approach chance performance, yet still leave identifiable behavioural traces that can be aggregated to improve detection.
By Qizhou Wang, Bogdan Mamaev, Christopher Leckie
The study investigates how pruning affects large language models (LLMs) used for smart‑home tool calling. Researchers examined four LLMs—dense Transformer, dense hybrid, and mixture‑of‑experts (MoE) architectures—using depth, width, hybrid, and expert pruning, followed by supervised fine‑tuning. They evaluated over 19,500 instances from three smart‑home datasets, analyzing not only overall accuracy but also degradation in action components (operation, device, argument, value) and task complexity, finding that dense models suffer sharp performance drops after a narrow safe pruning range, while MoE models tolerate more pruning; aggressive pruning also leads to over‑refusal and loss of grounded specificity.
By Congjing Zhang, Vashishtha Patil, Henning Lange, Usman Aleem
The paper investigates whether in‑context learning (ICL) can replace instruction tuning for multilingual language models, especially as model size varies. It highlights the difficulty of obtaining high‑quality instruction data in multilingual settings and compares the performance of ICL versus instruction‑tuned models. The findings show that a performance gap persists between the two approaches, suggesting the need for further research to close it.
By David Ponce, Thierry Etchegoyhen
PARSA‑Bench is the first dedicated benchmark for evaluating large audio‑language models on Persian, addressing unique challenges such as classical poetry, traditional music, and code‑switching. It comprises 16 tasks—10 of which are new—covering speech understanding, paralinguistic analysis, and culturally grounded audio reasoning. Across most tasks, text‑only baselines outperform audio‑based models, indicating that audio understanding remains the main limitation, except for Persian poetry where prosody provides additional information that audio beats text.
By Mohammad Javad Ranjbar Kalahroodi, Mohammad Amini, Parmis Bathayan, Heshaam Faili, Azadeh Shakery
YFPO (Yoked Feature Preference Optimization) is a neuron‑guided preference optimization framework that augments standard preference learning with internal neuron‑level rewards. It uses AttnLRP to identify math‑associated internal features and derives an auxiliary reward from the activation margin between preferred and dispreferred responses. Experiments on GSM8K with a compact language model show that these neuron‑guided rewards influence optimization dynamics and yield measurable improvements, indicating that internal representations can serve as lightweight, interpretable signals for reasoning‑oriented post‑training.
By Yifan Le
The paper explores whether CLIP embeddings can detect AI-generated images by using a frozen CLIP model to extract visual embeddings and training lightweight classifiers on top. On the CIFAKE benchmark, the approach achieves 95% accuracy without language reasoning, and 85% accuracy after few-shot adaptation with 20% of the data. Certain image types, such as wide-angle photographs and oil paintings, remain challenging, highlighting unexplored difficulties in AI-generated image classification.
By Ziyang Ou
R3 is a new reward modeling framework that addresses limitations in current reward models by being rubric‑agnostic and generalizable across multiple evaluation dimensions. It provides interpretable, reasoned score assignments rather than opaque scalar outputs, enhancing transparency and flexibility in evaluating language models. The authors release their models, data, and code openly at https://github.com/rubricreward/r3.
By David Anugraha, Zilu Tang, Lester James V. Miranda, Hanyang Zhao, Mohammad Rifqi Farhansyah, Garry Kuwanto, Derry Wijaya, Genta Indra Winata
The paper investigates whether large language models (LLMs) possess intrinsic value systems and how to quantify and align them. By projecting responses from 106 LLMs and 95,000 human survey profiles into a shared sociological space, the authors confirm that LLMs do have values, though these values form a concentrated, idealized core rather than mirroring human diversity. They introduce the Prior-Environment-Cognition (PEC) framework to mathematically define value expression and propose an adaptive Alignment Prescription that identifies minimal interventions—ranging from prompts to targeted parameter updates—to steer LLM values efficiently without harming general performance.
By Keqing Zhang, Jingyu Chen, Yufan Liu, Yongqiang Zhu, Nai Ding, Lai Jiang, Congyan Lang, Bing Li, Weiming Hu
The paper introduces ATRBench, a benchmark that measures the proactivity gap in long‑lived LLM agents by evaluating their ability to ask for user preferences that are not needed immediately but may be useful in future sessions. It defines the Ask‑to‑Remember (ATR) task, where agents must decide whether to request a reusable preference now, and shows that current state‑of‑the‑art agents perform significantly below an oracle. The study identifies preference acquisition as the main bottleneck and provides a diagnostic framework for improving agent proactivity.
By Bin Wu, Guanyun Zou, Bingbing Wang, Huan Zhao, Chuan Shi
arXiv:2408.14575v5 Announce Type: replace
Abstract: EVINCE (Entropy and Variation IN Conditional Exchanges) is a novel framework for optimizing multi-LLM dialogues using conditional statistics and in...
By Edward Y. Chang
arXiv:2609.16590v1 Announce Type: new
Abstract: People increasingly use frontier AI models for health advice, but via different access modes (e.g., ChatGPT, ChatGPT Health, APIs) with varying setting...
By Yuan Pu, Yewon Chang, Furong Jia, Xunjian Yin, Jessica Ma, Ayman Ali, Monica Agrawal
arXiv:2609.16058v1 Announce Type: cross
Abstract: Red-light violations and harsh braking at signalized intersections are major contributors to traffic accidents. This paper analyzes and predicts huma...
By Mohammad Khoshkdahan, Pavel Laskov, Alexey Vinel
arXiv:2609.16366v1 Announce Type: cross
Abstract: When foundation models describe people, recent work in AI fairness, accessibility, and ethics recommends avoiding inferred identity labels (e.g., "sh...
By Yingjia Wan, Lin Lin, Elisa Kreiss
arXiv:2609.16993v1 Announce Type: cross
Abstract: Large Language Models are now common in student assessment, but we know little about how student demographics affect their use. Sometimes, considerin...
By Donya Rooein, Luca Benedetto, Dirk Hovy
arXiv:2609.16179v1 Announce Type: new
Abstract: Z-loss has been widely applied to the logits of language-model output heads and sparse mixture-of-experts routers. Z-loss constrains the softmax log-no...
By Bum Jun Kim
arXiv:2609.16229v1 Announce Type: new
Abstract: Machine unlearning aims to remove specific knowledge from a trained large language model (LLM) without retraining from scratch. Existing methods modify...
By Pingzhi Li, Jinhao Duan, Vaishnav Tadiparthi, Nakul Agarwal, Kwonjoon Lee, Ehsan Moradi Pari, Hossein Nourkhiz Mahjoub, Sijia Liu, Tianlong Chen