WeAgent-MMSearch introduces a multimodal search agent that preserves retrieved images as persistent references, enabling the model to inspect, process, and cite them throughout a search trajectory. The system includes a harness (WeAgent-Harness), a post‑training method (FA‑GSPO) that recovers salvageable rollouts, and a new benchmark (VisTarget‑Bench) to evaluate image‑retrieval versus visual‑perception failures. Evaluation shows that agentic post‑training boosts performance by 19.22 points, allowing the model to outperform similarly sized open‑source models and compete with much larger ones.
By Zongkai Liu, Hui Zhang, Liqiang Niu, Zhen Cao, Han Li, Juntao Liu, Wenchao Chen, Chengduo Zhao, Chao Yu, Fandong Meng
RoboPhys-3D is a 3D‑grounded embodied world model benchmark built on RoboTwin 2.0, featuring 50 manipulation tasks, 5,000 episodes, and 25,000 multi‑view ground‑truth videos. It evaluates video world models by processing both generated and ground‑truth videos through the same 3D reconstruction pipeline, allowing the separation of reconstruction‑induced from generation‑induced errors. The benchmark defines 50 metrics across four sub‑dimensions—pixel fidelity, 3D geometry consistency, state understanding, and task completeness—and introduces the Average Full Score and RoboPhyscore for holistic assessment, with RoboPhyscore showing strong correlation with human judgments.
By Tianyi Wang, Jiazhou Chen, Yiming Xu, Xiangyu Li, Tianyi Zeng, Chih-Hsien Chou, Ning Lu, Liang Peng, Junfeng Jiao, Christian Claudel
WorldProbe is a new diagnostic benchmark that tests video world models by letting a goal‑conditioned multimodal agent adapt camera actions online during generation, creating 164 test cases across four probing families. The benchmark evaluates 10 state‑of‑the‑art models over more than 1,600 rollouts, totaling roughly 800 minutes and horizons up to 180 seconds, revealing that while local visual and content quality can be maintained for longer, long‑range memory and evolution remain weak. The study highlights a gap between locally stable generation and persistent world consistency that existing short‑interaction benchmarks miss.
arXiv:2609.15478v1 Announce Type: new
Abstract: Multimodal agents can create complex videos in software such as Blender by coding without relying on diffusion models. Yet video understanding benchmar...
By Yolo Y. Tang, Daiki Shimada, Jiayue Meng, Jing Bi, Pinxin Liu, Yicheng Wang, Yunzhong Xiao, Zhangyun Tan, Zeliang Zhang, Chao Huang, Susan Liang, Qianxiang Shen, Luchuan Song, Ali Vosoughi, Mingqian Feng, Melika Filvantorkaman, Chenliang Xu
arXiv:2608. 03979v1 Announce Type: cross Abstract: We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration.
By Zhen Fang, Yu Zeng, Wenxuan Huang, Yiming Zhao, Shiting Huang, Tianfei Ren, Qi Lu, Qingnan Ren, Qisheng Su, Lionel Z. Wang, Qingyu Yin, Shuang Chen, Zehui Chen, Lin Chen, Zhenfei Yin, Yao Hu, Shaohui Lin, Wanli Ouyang, Shaosheng Cao, Feng Zhao
arXiv:2609.21257v1 Announce Type: cross
Abstract: Large language model (LLM) agents can propose, implement, and evaluate model changes. Autoresearch loops demonstrate this capability through minutes-...
By Hao Fu, Baiting Zhu, Minglei Chen, Yinjie Huang, Shuai Ding
AdaVDR is an adaptive video deep research agent that selects and reflects on tool usage based on the task and the model’s capabilities. It constructs a specialized data pipeline to generate high‑quality QA pairs and uses model‑conditioned filtering to remove unnecessary tool calls. The agent is trained with supervised fine‑tuning and reinforcement learning, achieving top performance on the VDR‑EE benchmark and significant gains on VideoDR.
By Xintong Zhang, Xiaomeng Fan, Shilin Yan, Ekko He, Zicheng Liu, Zijian Zou, Guannan Zhang, Yuwei Wu, Zhi Gao, Hongwei Xue
arXiv:2512.23044v3 Announce Type: replace
Abstract: Autonomous agents are shifting information seeking from passive retrieval toward proactive, open-ended research, yet this progress has largely bypa...
By Zhengyang Liang, Yan Shu, Xiangrui Liu, Minghao Qin, Nicu Sebe, Jiaxin Mao, Zheng Liu, Lizi Liao
arXiv:2608.23329v1 Announce Type: cross
Abstract: Open-world video understanding often requires a model to locate sparse visual evidence and acquire external knowledge that is absent from the video a...
By Wenqi Liu, Shijie Ma, Yunxiao Wang, Meng Liu, Qile Su, Han Liu, Bohan Hou, Xuanyu Zheng, Changyi Liu, Tianke Zhang, Haonan Fan, Kaiyu Jiang, Yingxin Li, Jiankang Chen, Xu Wang, Bin Wen, Tingting Gao, Han Li, Jianhua Yin, Yinwei Wei, Xuemeng Song
The paper introduces VHOP, a data generation framework and benchmark for visual agentic search, and VHOP-Router, an end‑to‑end training pipeline that turns a standard embedding model into an autoregressive multi‑step retriever operating directly in visual latent space. VHOP-Router eliminates the need for intermediate text queries, boosting retrieval accuracy from under 5% to 76.3% and improving task success rates by 52.7% while dramatically reducing token usage and API payloads. The approach generalizes to unseen difficulty levels and realistic test sets, offering an efficient solution that preserves native LLM capabilities.
By Tianyu Chen, Mingyuan Zhou, Jiaxing Wu
arXiv:2610.04432v2 Announce Type: replace
Abstract: Building interactive simulators from real-world observations is a promising way to scale embodied data, but current pipelines still rely heavily on...
By Jinzhou Tang, Zijun Zhang, Jing Yang, Yuchen Yan, Kun Zhou, Lingjun Mao, Ruobing Han, Jinglin Cao, Wenpeng Xu, Lukun He, Minghao Fu, Fan Feng, Biwei Huang
AgentVQA is an offline benchmark designed to evaluate vision‑language models on realistic agentic scenarios across five domains: Web Agents, Robotics, Egocentric Videos, Games, and Spatial Planning. It comprises 1,000 carefully selected real‑world examples drawn from 23 source datasets, enabling reliable comparisons without costly online simulations. The benchmark reveals that current models achieve only 58.5% accuracy, with most errors stemming from difficulties in grounding and tracking evolving visual states rather than from task misunderstanding or reasoning.
By Advait Gupta, Sachin Konan, Tom Yin, Gabriel Sarch, Zhuang Liu