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
science-buddy
cs.AI, cs.CL
2026-09-16
ScienceBuddy nests frozen-model harness search inside GRPO on Qwen3.5-4B. Three cycles raise single-attempt accuracy from 42.2% to 73.3% on four biomedical task families.
Scientific agents already retrieve papers, query databases, and run analysis scripts. A researcher correcting one answer in chat still does not make the next task go better. Reflexion, GEPA, and ACE rewrite prompts and playbooks. SIA and HELIX start moving both the harness and the weights. What is missing is a closed loop that turns everyday collaboration into executable tasks and scoring rubrics, then lets working procedures and the model feed each other.
ScienceBuddy is an interactive scientific workspace released by PhAI Labs. The tool catalog comes from Biomni: 224 tools across 22 modules, covering genomics, molecular and cancer biology, pharmacology, bioimaging, literature, and databases. Corresponding authors sit at Oxford, Stanford, and Princeton. The product puts researchers in the workflow and turns their requests, feedback, and execution traces into continual-learning fuel.
The mechanism is recursive-in-recursive self-improvement, two nested loops.
The inner loop freezes the task model. A separate, fixed helper, GPT-6 Astra, reads failed trajectories and proposes one bounded edit: one instruction, one scoped skill, or one context setting. A candidate is kept only if it passes edit constraints and raises the mean rubric score on a paired development set. Ties keep the parent. Tools, runtime, and evaluators stay frozen.
The outer loop freezes the selected harness, calibrates and augments task environments against current ability, then runs GRPO on fresh on-policy rollouts with task-specific rubric rewards. Historical chats define tasks and diagnoses; they are not on-policy training samples. Researcher replies are not treated as ground truth.
The two loops feed each other. Change the harness, and the training trajectories and difficulty move with it. Change the model, and the same procedures start to misfit. At the start of cycle two, harness validation drops from 44.4% to 34.4%, which is the coupling showing up. The live service keeps serving the current (model, harness) pair; background updates ship together. Tasks are packed as Harbor bundles: instruction, assets, and rubric checks in tests.
The run starts from Qwen3.5-4B plus an initial scientific harness. The pool is 895 items from LAB-Bench and Biomni-Eval1: LitQA2 96, DbQA 511, ProtocolQA 108, GWAS 180.
Three co-evolution cycles, 10 inner search steps and 20 GRPO updates each. Single-attempt test accuracy moves from 42.2% to 73.3%. 33.3% of items flip from wrong to right; 2.2% go the other way. Approximate family numbers from the figure: DbQA 36% to 74%, GWAS 30% to 60%, LitQA2 80% to 87.5%, ProtocolQA 60% to 70%. Harness validation by cycle: 38.9% to 44.4%, 34.4% to 46.7%, 61.1% to 70.0%. Mean RL reward in the first versus second half of each outer stage: 33.3% to 38.8%, 44.1% to 60.5%, 57.8% to 69.8%.
Each mechanism also works alone. Freeze the model and edit the harness: first-response validation accuracy 31.1% to 51.1%, plus 20 points. The selected harness holds four instructions and nine scoped skills. Freeze the initial harness and train for about two hours: pass@4 coverage 48.3% to 67.8%, plus 19.5 points.
For anyone building a science agent, this writes "edit the playbook" and "train the model" as an alternating engineering loop, and ships the workspace, task format, and rubrics. A 4B model gaining 31 points on tool-rich biomedical QA says the procedures and skill text are doing a lot of the work.
This is a systems and product paper, not a new optimizer. The inner editor stays a frozen GPT-6 Astra. The tasks remain lab questions with reference answers. Open-ended discovery is still a slogan.
The paper is explicit: the reflector stays fixed, so better task scores do not mean the improvement machinery itself got stronger. User feedback in the harness case comes from a Qwen3.8-27B simulator over 288 adaptation conversations; the allowed replies never contain the correct answer. Real researcher traces are two qualitative cases, JAK1 study design and an ARL4C talk, and Figure 7 labels the derived specs as illustrations, not archived rubric packages.
Gains are not attributed to individual skills or feedback sources. Evaluation is almost entirely biomedical multiple-choice and retrieval. The workspace claims multiple domains; the tools and data still skew biomedical. The 42% to 73% jump is against the same 4B starting point. A full coupled ablation that spends the same compute on GRPO only, no harness search, is not in the main figure.