科學代理／持續學習
ScienceBuddy Alternately Rewrites the Agent Harness and Model, Turning Research Interactions into Ongoing Training Material
ScienceBuddy converts researchers’ requests, feedback, and execution logs into tasks and scoring rules, then alternately optimizes the agent code and model. The team has released a simplified experimental framework, but the full code for the hosted workspace and evidence of long-term self-improvement remain unpublished.
PhAI Labs and its collaborators have released ScienceBuddy, an effort to transform scientific agents from one-off question-answering tools into systems that continually adapt alongside research workflows. Its workspace can ingest papers, spreadsheets, biological sequences, and scientific images while retaining plans, tool inputs and outputs, and analytical artifacts. The version described in the paper covers 22 modules and 224 tools, focused primarily on genomics, molecular and cancer biology, pharmacology, imaging, and literature retrieval.
The core method is called “recursion within recursion.” In the inner loop, the model is held fixed while interactions and validation results are used to propose new Python harnesses, with candidate versions competing against the parent version on a fixed validation set. In the outer loop, the selected harness is held fixed while the task model is updated via SkyRL GRPO using verifier rewards. The new model then enters the next round of harness search, so improvements affect not only the weights but also the tools, procedures, and scoring processes available to the model.
The public experiment uses Qwen3.5-4B, with the data divided into 715 training, 90 validation, and 90 test tasks. It runs three harness/RL cycles. Each harness stage consists of only three steps, with 16 interactions and three candidates per step, while the RL stage performs 30 GRPO updates. This allows researchers to inspect how state is handed off between the two nested loops instead of seeing only a product demonstration.
The limitations are equally important. The open-source repository contains only Simple-SciBuddy, the execution agent, verifiers, and the training interface; it does not include the hosted workspace frontend, account system, or full product API. The experiments use frozen data and simulated feedback, and they do not yet demonstrate that the system can learn safely over the long term in real laboratories. If unpublished data and researcher feedback are incorporated into training in practice, engineering teams will also need to clarify data ownership, contamination, rollback, and scorer drift. The next things to watch are complete benchmark results, cross-domain replication, and whether every self-modification produces an auditable version.