KnowDemo is a framework that generates diverse robot demonstrations from human videos by leveraging structured manipulation knowledge. It uses a vision‑language model to extract task requirements and permissible execution variations, then resolves these against target‑scene entities to guide candidate generation and screening before motion planning. The resulting demonstrations feature multimodal behavior, alternative contact strategies, and valid subtask orders, and have been shown to improve planning success and enable sim‑to‑real policy transfer across three tasks.
By Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Sch\"afer, Michael Beetz
CESBench is a new benchmark for evaluating large language models on cryptographic engineering security for IoT devices, comprising 380 expert‑written items across six sub‑domains such as side‑channel, fault injection, and implementation. The benchmark includes four task types—multiple‑choice, judgment, scenario, and code—each designed to test different competences, with automatic scoring for the first two and LLM‑based judging for the latter two. Evaluation of 11 open‑weight and proprietary LLMs shows strong performance on multiple‑choice and code tasks but weaker results on judgment and scenario tasks, highlighting gaps in justifying security verdicts.
By Wenquan Zhou, An Wang, Jing Liang, Peien Feng, Jingqi Zhang, Yaoling Ding, Liehuang Zhu
The paper introduces a Situation–Internal state–Behavior Persona method to improve large language models’ ability to impersonate real individuals in social media contexts. It also proposes an evaluation protocol that supplies LLM evaluators with reference information about the target individual. Experiments on a new dataset of social media replies show the method surpasses state‑of‑the‑art in‑context learning baselines, and the protocol correlates moderately with human judgments, while additional tests on fictional characters confirm broader applicability.
By Ji-Lun Peng, Yi-Zhen Zhang, Chun-Nan Chou, Yun-Nung Chen
AgentVidBench is a new multi‑hop video question‑answering benchmark designed to evaluate spatial, temporal, and causal reasoning in multimodal large language models (MLLMs). Unlike existing tests that focus on simple scene queries or global summaries, AgentVidBench includes step‑by‑step solution traces to assess whether agents gather the necessary evidence to justify their answers. Experiments with 12 MLLMs show limited single‑turn performance, but integrating these models into agentic workflows improves both accuracy and trajectory scores, establishing AgentVidBench as a comprehensive testbed for future research on agentic video understanding.
By Seoyeon An, Hyeonseo Jang, Minsu Kim, Chanho Lee, Younghan Park, Kangwook Lee
LoG-VGGT is a memory‑efficient framework for long‑sequence 3D reconstruction that balances local temporal modeling with global camera consistency. It uses cross‑window attention in a small subset of transformer blocks to propagate information across adjacent temporal windows while keeping memory usage bounded. A global camera consistency refinement module further improves long‑horizon pose stability by enforcing scene‑level constraints through cross‑attention between camera and compact register tokens, leading to better depth accuracy and robust camera pose estimation on multiple benchmarks.
By Jingke Zhou, Chenhang Ma, Zhizhou Zhong, Mingkai Liu, Zhuang Zhou, Yicheng ji, Binghua Su, Bo Cai, Xianliang Huang
HE-Guardrail is a framework that applies homomorphic encryption to enforce guardrails against jailbreak attacks during encrypted large language model inference. It evaluates guardrail mechanisms—Llama Guard, JBShield, and GradSafe—directly on encrypted data, deciding whether to return the model’s response to the client. The approach preserves confidentiality while closely matching the decisions of plaintext guardrails, offering varied security-efficiency-utility trade‑offs.
By Byeongseo Min, Yongwoo Lee, Young-Sik Kim, Yongjune Kim
VidOmni-Bench is a new benchmark for fine‑grained video understanding that asks models to verify whether each event in dense video captions is supported by the video. It contains 500 videos covering five complexity types and durations from 4 seconds to 90 minutes, and uses human‑verified sentence‑level labels to create hard negatives. Experiments show that Video‑LLMs often hallucinate events, struggle to detect incorrect descriptions, and exhibit varying weaknesses depending on video complexity and duration.
By Changbeen Kim, Junwon Chang, Kipyo Kim, Risa Shinoda, Kuniaki Saito, Donghyun Kim
OneBid is a unified auto‑bidding foundation model that consolidates diverse cost‑per‑X (oCPX) advertising scenarios into a single framework. It builds on Decision Transformer by conditioning on two atomic signals—Return‑to‑Go for conversion value and Cost‑to‑Go for cost ratio—and incorporates value‑aware regularization. A sequence‑level Mixture‑of‑Experts architecture captures cross‑scenario knowledge while preserving low latency, and a Critic‑guided Relative Offline Policy optimization (CROP) aligns the backbone with scenario‑specific preferences without unsafe online exploration. In production at Kuaishou, OneBid achieved a 2.2% overall ADVV increase and up to 13.1% in the ROAS scenario.
By Yewen Li, Peng Jiang, Yitian Li, Pengfei Lv, Xialong Liu, Peng Jiang, Qingpeng Cai
SynthDemo‑RL introduces a teacher‑student framework that uses an automated teacher to generate successful manipulation trajectories from simulator‑privileged state, which are then distilled into a Vision‑Language‑Action (VLA) student via supervised fine‑tuning. The student is further refined with PPO using binary task‑success rewards. On the LIBERO‑PRO benchmark, SynthDemo‑RL rescues all 27 previously unsolvable tasks and achieves near‑perfect success rates, matching performance that would otherwise require human demonstrations.
By Hiroaki Kingetsu, Hiroaki Kurihara, Kaoru Yokoo, Kenji Fukumizu, Manohar Kaul
The paper investigates the effectiveness of activation steering in latent chain-of-thought (CoT) reasoning compared to explicit CoT. It finds that steering continuous latent thoughts yields weaker impacts on language generation, even when hidden representations are shifted similarly. The authors propose a latent-to-language transition gap, supported by evidence of abrupt output distribution changes at the transition boundary and weaker bidirectional control in latent CoT.
By Gaoxiang Huang, Lei Qi
CIBuzzBench is a new benchmark that tests cross‑lingual understanding of Chinese internet buzzwords by providing 3,001 buzzwords with English explanations, equivalents, category labels, and harmfulness annotations. The benchmark defines three tasks—Meaning Explanation, Cross‑lingual Equivalent Matching, and Culturally Grounded Harmfulness Detection—to evaluate how well large language models translate and interpret these culturally nuanced terms. Experiments show that current state‑of‑the‑art LLMs still struggle with fine‑grained non‑literal meanings, robust matching, and accurate harmfulness detection across languages.
By Yifan Wang, Junyu Lu, Qifan Wang, Shun Zhang, Chaozhuo Li, Jiahao Liu, Zhijun Cao, Lingbin Bu, Fanliang Bu
The paper examines whether fine‑tuning large language models (LLMs) with personality‑labelled data improves their ability to act as socially interactive agents. Two small open‑weight LLMs were fine‑tuned on a corpus of personality‑labelled social media posts and dialogues, and the resulting models were evaluated in various social interaction scenarios by independent LLM judges. The findings show that the fine‑tuned models do not outperform their baseline counterparts in role‑playing personalities, though they offer comparable text quality and increased linguistic diversity for the Qwen models; low inter‑rater agreement limits confidence in the results, suggesting future work should focus on training data quality and domain alignment.
By Tim Krabbe, Xiaodan Shi
arXiv:2609.21888v1 Announce Type: cross
Abstract: Detecting pretraining data in large language models is challenging because high likelihood can reflect either training exposure or strong generalizat...
By Chenye Ke, Zirui Liu, Qi Liu, Yan Zhuang, Jintao Zhang, Zhenya Huang, Shijin Wang
arXiv:2609.21942v1 Announce Type: cross
Abstract: A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt...
By Eshika Pathak, Leela Krishna
The paper examines how users delegate tasks to the AI agent OpenClaw by analyzing 73,093 Reddit posts. It identifies 21 human values grouped into six categories—such as Autonomous Operation, Dependable Operation, Affordable Operation, Bounded Reach, Reviewability, and Equitable Access—and finds that values are largely satisfied when users describe the agent’s outputs but often unmet when users discuss supervising the agent. The authors term this pattern "value‑sensitive delegation," emphasizing that supporting human values requires attention to both what an agent does and the conditions users set around its use.
By Renkai Ma, Ruyuan Wan, Xuan Lu, Fan Yang, Chen Chen, Lingyao Li
MemeLens is a unified multilingual, multitask Vision‑Language Model designed to improve meme understanding across a wide range of tasks such as hate, misogyny, propaganda, sentiment, and humour. The authors consolidated 38 public meme datasets, mapping their labels into a shared taxonomy of 20 tasks covering harm, targets, figurative intent, and affect, and conducted extensive experiments to show that multimodal training and a unified approach outperform fine‑tuning on individual datasets. All experimental resources, the model, and the datasets are released publicly for community use.
By Ali Ezzat Shahroor, Mohamed Bayan Kmainasi, Abul Hasnat, Dimitar Dimitrov, Giovanni Da San Martino, Preslav Nakov, Firoj Alam
arXiv:2609. 21362v1 Announce Type: new Abstract: Conventional tokenizers represent text as characters or statistically derived subwords, overlooking the internal phonological structure of syllables and often requiring large vocabularies.
By Nghia Hieu Nguyen, Thai Bao Huynh, Binh-An Dinh-Le, Phu Gia Hoang, Dat Tien Nguyen, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen
arXiv:2609.14744v2 Announce Type: replace
Abstract: By acquiring compute, credentials, accounts, services, and other agents, autonomous AI agents can introduce new authority into a task. Payment, bud...
By Genliang Zhu (Accentrust, Georgia Institute of Technology), Chu Wang (Accentrust, University of Illinois Urbana-Champaign)
The paper investigates how transformer language models perform few‑shot learning for a simple addition task, showing that the ability is concentrated in a handful of attention heads. Using dimensionality reduction, the authors identify low‑dimensional subspaces—three heads with six‑dimensional spaces in Llama‑3‑8B‑Instruct—where specific dimensions encode the units digit via trigonometric patterns and magnitude via low‑frequency components. They also derive a mathematical identity linking aggregator and extractor subspaces, enabling tracking of information flow from examples to the final prediction.
By Xinyan Hu, Kayo Yin, Michael I. Jordan, Jacob Steinhardt, Lijie Chen
The paper introduces a chemical language foundation model that uses a SMILES‑based polymer graph representation (CPG) to encode polymer architecture and connectivity, addressing gaps in existing line notations. The model achieves strong performance across 30 polymer property benchmarks and demonstrates robustness to structural representation perturbations, with even chemically invalid SMILES sometimes matching state‑of‑the‑art results. Control experiments and attention analyses confirm that CPG offers meaningful advantages while highlighting the model’s ability to interpolate SMILES sequence space in a way that aligns loosely with chemical and architectural space.
By Nathaniel H. Park, Eduardo Soares, Victor Y. Shirasuna, Tiffany J. Callahan, Sara Capponi, Emilio Vital Brazil