01arXiv · 2026arXiv · 2026
An agentic simulation-ready world pipeline connecting ISArt assets, Unravel reconstruction, SimForge text-to-scene generation, validation, and PolicyForge demonstrations.
连接 ISArt 资产、Unravel 重建、SimForge 文本生成场景、验证与 PolicyForge 示范数据的智能体仿真世界系统。
02ICLR 2027 submissionICLR 2027 投稿
RSI-CUA
A Physical RSI framework for computer-use agents: environments, rollouts, feedback, and iterative improvement in an endless training ground.
面向计算机使用智能体的 Physical RSI 框架:在持续训练环境中连接环境、rollout、反馈与迭代改进。
03Findings of EMNLP 2026EMNLP 2026 Findings
A temporal scientific knowledge-graph system for reconstructing research evolution, grounding verification in evidence, and supporting trend forecasting.
面向科学研究演化重建的时序知识图谱系统,用证据支撑科学验证,并辅助趋势预测。
04EMNLP 2026 Oral · Outstanding Paper AwardEMNLP 2026 Oral · Outstanding Paper Award
A perception–reasoning benchmark for medical-image quality across 5 modalities and 40+ attributes, with 2,600 perception queries and 708 reasoning assessments.
面向医学影像质量的感知—推理评测基准,覆盖 5 种模态、40 多项属性、2,600 个感知问题与 708 个推理评估。
05arXiv · 2024
A multi-module LLM agent for patent QA, image-to-molecule extraction, and central-compound identification.
面向专利问答、结构图像转分子与核心化合物识别的多模块大模型智能体。
06arXiv · 2026
A diagnostic benchmark for reliable R-group grounding in Markush molecular editing.
面向 Markush 分子编辑中可靠 R 基团 grounding 的诊断型评测基准。