From the source · Translated from Chinese
Recently, Shanghai AI Laboratory launched the general-specialty integrated agent Intern-A1.5 (Sage-A1.5), advancing recursive self-improvement (RSI) research further. Compared to version 1.0, Sage-A1.5 further enhances agent orchestration, autonomous verification, and complex task-solving capabilities. It builds on the long-horizon task capabilities of the precursor model Agents-A1, relying on the Intern·Duanyan Scientific Discovery Platform, and validates the RSI flywheel in scenarios such as neuroscience and electrochemistry—being the first to verify the feasibility of a base model self-evolution technical path driven by real scientific scenarios. This path is a practical implementation of the 'environment as data' concept proposed by Director Zhou Bowen—the agent continuously accumulates domain knowledge and corrects reasoning deviations in the real scientific research process, driving the simultaneous iteration of general foundational capabilities and specialized scientific capabilities, and opening an internal closed loop for autonomous scientific exploration and capability upgrading. Experience link: https://discovery.intern-ai.org.cn/token-plan/home?tabIndex=0. Intelligence capability advancement: the agent can collaborate and verify, supporting complex scientific exploration. Research and engineering tasks often require multiple steps such as retrieving information, proposing hypotheses, writing code, and verifying results. The longer the task, the more reasonable division of labor and timely error correction are needed. Based on the Intern·Duanyan Scientific Discovery Platform, the team incorporated the organization and adjustment execution process of scientific research task workflows into the Sage-A1.5 training process, expanding the model's capability boundaries while training the model to combine existing capabilities and dynamically adjust subsequent actions based on new evidence. For the challenges of multi-thread exploration and multi-source evidence integration in complex research, the team built Agentic Orchestration capability. The trained model can autonomously decompose tasks, dispatch sub-agents for parallel investigation, and summarize results. Data shows that after enabling this capability, Sage-A1.5's scores on long-horizon retrieval and multi-evidence comprehensive report evaluations all improved. In a space science review case, Sage-A1.5 dispatched 18 sub-agents, checked 24 assertions against 12 space missions, and compared evidence and competing explanations in parallel around six methodological challenges. It identified the logical leap of 'detecting organic molecules proves the origin of life,' investigated three small-body sampling missions, and distinguished the evidence requirements for molecular existence, extraterrestrial origin, and biological origin layer by layer; for the K2-18 b spectral controversy, it autonomously compared multiple research plans, sorted out pros and cons, reviewed original literature, and organized the results into a checklist, evidence matrix, and revision record to support the review revision. To address the problems of repeated retrieval and error accumulation in long-horizon research, the team built Agentic Verification capability. The trained model can autonomously determine the timing and object of verification and advance exploration based on verification results. When research encounters obstacles or key conclusions need confirmation, the model can distinguish confirmed facts from hypotheses to be verified, identify evidence gaps and failure paths, supplement evidence, correct judgments, or adjust strategies, all triggered autonomously without relying on external preset rules. Taking a photochemical reaction study as an example.
shlab.org.cn
