Technical reports and papers

1. 15 Sep 2026Foundation Models

   Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

   Ling Yang, Zhenfei Yin, Yingcheng Wu

   Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and acting within problems specified by humans to participating in the process by which new problems, representations, explanations, and knowledge are created. We refer to this capability as Discovery Intelligence. We formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. A DFM operates over a revisable research state and supports seven coupled capabilities spanning problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement. We instantiate this framework with Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. We further ground the framework with GALILEO, a real therapeutic-discovery system in which Dry-Lab reasoning, robotic and hands-on Wet-Lab experimentation, external biological evidence, and iterative hypothesis and design revision form a closed physical discovery loop. We then formulate a unified approach to capability formation and process-centered evaluation, enabling discovery behavior to be trained, improved, and measured beyond final-answer performance. Together, these components establish discovery as a learnable, executable, and evaluable capability of foundation-model systems. We view this shift as a broader progression in intelligence scaling: from learning over existing knowledge, to learning from action outcomes, and ultimately to participating in the construction, testing, and revision of the structures through which new knowledge is discovered.

   PhAI Labs Technical Report

2. 16 Sep 2026Scientific Agents

   ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents

   Shuhan Xue*, Jianyuan Zhong*, Ziyuan Nan*, Wenbin Li, Zhaochen Yu, Jinchao Ding, Qiang Gao, Pengyu Zhan, Yuntong Zhang, Tian Cheng, Zhenfei Yin, Yingcheng Wu, Ling Yang

   ScienceBuddy is an interactive scientific research workspace that brings continually improving agents into a researcher's everyday work. It supports researchers in carrying out scientific tasks while turning their requests, feedback and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, which couples harness evolution with model reinforcement learning: the inner recursion improves the harness with the model fixed, and the outer recursion trains the model under the improved harness. Case studies cover researcher interaction, harness refinement and model learning across four families of scientific task.

   PhAI Labs Technical ReportReleasing 16 Sept

3. 17 Sep 2026Research Environments

   ScienceIDE: Scaling Scientific Experience toward 1,000 Executable Environments

   Hejia Geng*, Zesen Huang*, Haoyang Li, Wenbin Li, Koutian Wu, Zihan Zhou, Yuanbo Pang, Weihao Liu, Zigong Xu, Yueheng Shi, Yue Ma, Zongzheng Zhang, Tianzhe Zheng, Chuanfei Dong, Fengyu Xie, Zhiping Li, Jiankai Sun, Qucheng Gao, Jiaming Pan, Zhenlin Zhu, Peijin Zhang, Lanqing Yuan, Liuwei Xu, Xianrong Liu, Tong Xie, Junkai Wang, Zonglin Di, Xiao-Han Ma, Kangkai Liang, Ziang Liu, Sheng Huang, Zehong Zhao, Ziyang Xu, Jingxu Xie, Yaopeng Xing, Jiayi Xian, Xing Meng, Zhenfei Yin, Yingcheng Wu, Ling Yang

   Scientific software and expertise do not readily become experience that agents can learn from: repositories depend on fragile toolchains and unwritten conventions, and correctness rests on scientific behaviour and numerical tolerances. ScienceIDE is infrastructure that connects scientific practice with agent development by making expert knowledge reusable. Domain experts define scientific cases and acceptance criteria, agents help turn those decisions into executable environments, and task factories generate diverse challenges whose validity is established through execution and scientific verification. The same environments yield graded trajectories for supervised fine-tuning, online rewards for reinforcement learning, and evidence for evaluation.

   PhAI Labs Technical ReportReleasing 17 Sept

4. 18 Sep 2026World Models

   JEPA-Anything: Learning Predictive Models across Different Worlds

   World models learn internal states for predicting how a system changes. Cells, molecules, physical fields, control environments and clinical trajectories differ radically, yet share one predictive problem: infer unobserved, intervened or future states from the current context. JEPA-Anything is a domain-agnostic framework that builds on joint-embedding predictive architectures and replaces their monolithic target representation with orthogonal predictive factorization, which organises a latent world state into complementary predictive factors, learns each through its own pathway, and recombines them for readout, intervention prediction and rollout. Each domain keeps its own observations and encoders while sharing the predictive core.

   PhAI Labs Technical ReportReleasing 18 Sept

4 in total