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, et al.
cs.AI · cs.CL
Quick verdict
Couples harness evolution with rubric-reward RL for a scientific agent (42.2→73.3% over 3 cycles); promising, but evaluation is small, single-run, and simulated-user based.
Summary
ScienceBuddy is a released interactive workspace for biomedical research assistance that aims to convert researcher collaboration into continual agent improvement. Its 'recursive-in-recursive' paradigm nests two loops: an inner loop in which a fixed editor model (GPT-6 Astra) diagnoses trajectories and proposes bounded instruction/skill edits to the harness, accepted only if valid and improving paired development rubric scores; and an outer loop that trains the task model (Qwen3.5-4B) with GRPO using task-specific rubric rewards on fresh on-policy rollouts under the selected harness, with environment difficulty calibrated by pilot execution. Interactions are packaged as Harbor tasks with rubrics, explicitly not treating user approval as ground truth. Four case studies: qualitative real interactions (JAK1, ARL4C); three coupled cycles raising held-out single-attempt accuracy from 42.2% to 73.3% on 895 LAB-Bench/Biomni-Eval1-derived tasks (33.3% newly solved, 2.2% regressed); harness-only adaptation raising validation first-response accuracy 31.1%→51.1% at fixed weights; RL-only raising pass@4 coverage 48.3%→67.8%. Main limitations: small evaluation panels with no confidence intervals or significance tests; quantitative adaptation uses a simulated reference-assisted user rather than real researchers; no equal-compute baseline isolating the value of coupling versus harness-only or RL-only; LLM-judged rubric rewards lack human agreement checks; and Section 4.2 is titled 'Two-Cycle' while describing three cycles.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Researcher collaboration (requests, feedback, execution evidence) can be converted into executable Harbor tasks plus task-specific rubrics that supervise post-training without treating user approval as scientific ground truth.
- Feedback-guided harness evolution with a fixed task model, gated by paired development evaluation, improves first-response validation accuracy by 20 percentage points (31.1%→51.1%) without any weight updates.
- Continual RL with task-adaptive rubric rewards under a fixed harness expands pass@4 problem coverage from 48.3% to 67.8% on the same evaluation budget.
- Alternating the two recursions across three cycles raises held-out single-attempt accuracy from 42.2% to 73.3% with gains in all four task families, while background updates keep the online service continuously available.
Strengths
Clean separation of learning signals from user feedback
Table S2 defines explicit information boundaries per role; the diagnostic feedback score q_t is kept out of policy optimization — GRPO uses only the separately evaluated trajectory reward R_x(τ) (Section S1.3) — and rubrics are fixed during optimization and paired harness evaluation (Sections 2.3–2.4), avoiding the common leakage of user approval into rewards.
Gated, regression-aware harness evolution
Equation 5 accepts a candidate only if it passes schema validation and improves mean normalized rubric score on identical paired development tasks, with previously successful tasks included to catch regressions and ties retaining the parent (Section 2.3); edits are bounded to one scoped skill/instruction change with a size budget.
Component-isolated evidence with honest self-scoping
Evaluation grounded in established public benchmarks
Product-level engineering with operational continuity
Weaknesses
No uncertainty quantification on any headline result
Quantitative experiments use simulated, not real, researcher feedback
No baseline establishing the value of coupling
Opaque judge and editor; reconstructed prompts hurt reproducibility
Internal inconsistencies and underexplained dynamics
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is 'recursive-in-recursive' self-improvement in ScienceBuddy?
It is a two-loop training scheme. The inner recursion improves the agent's harness — its instructions, skills, and context settings — while the language model is frozen: an editor model diagnoses failure trajectories and proposes small edits, which are kept only if they raise rubric scores on paired development tasks. The outer recursion then trains the model itself with reinforcement learning (GRPO) using rubric-derived rewards, while the improved harness is held fixed. The two loops alternate, each cycle's output feeding the next.
How much did ScienceBuddy improve its underlying model?
The paper reports three case-study results on 895 biology tasks derived from LAB-Bench and Biomni-Eval1: harness-only editing (model weights fixed) raised validation first-response accuracy from 31.1% to 51.1%; model-only RL (harness fixed) raised pass@4 problem coverage from 48.3% to 67.8%; and three coupled cycles raised held-out single-attempt accuracy from 42.2% to 73.3%. All numbers come from single runs without confidence intervals.
Was ScienceBuddy trained on real scientists' feedback?
Only qualitatively. The paper shows two real researcher interactions (a JAK1 study-design request and an ARL4C presentation task) to illustrate how requests become tasks and rubrics, but the quantitative harness-adaptation experiment used a simulated user — a smaller auxiliary model that issued bounded, pre-defined replies guided by reference correctness — rather than human researchers.
Is ScienceBuddy's code and data available?
The paper links a website (science-buddy.io) and a code repository (Gen-Verse/ScienceBuddy on GitHub) and releases the workspace as a product. However, the role prompts in the appendix are described as expository reconstructions rather than archived payloads, the RL hyperparameters are given only symbolically, and the harness editor depends on a proprietary model (GPT-6 Astra), so full independent replication is not straightforward.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure data is flattened in extraction: the per-family before/after pairing in Figure 8(c) is ambiguous from the text, so this review does not assert per-family numbers.
- The paper text appears complete (main sections plus appendix), but UI screenshots and recording-based figures (Figures 4–6, S2–S3) cannot be verified from text; the paper itself notes dialogue is reconstructed from recordings.
- Code release (Gen-Verse/ScienceBuddy) and the science-buddy.io website are as claimed in the paper and were not independently verified.
- The metadata author list (9 names) is shorter than the author list in the paper text (15 names); the review follows the paper text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.