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

ScienceBuddy is a newly released interactive scientific research workspace whose recursive-in-recursive self-improvement paradigm couples harness evolution (inner recursion) with model reinforcement learning (outer recursion), enabling scientific agents to…

The authors release ScienceBuddy, an interactive scientific research workspace that transforms researcher requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its core contribution is a recursive-in-recursive self-improvement paradigm: the inner recursion evolves the evaluation harness while the model is held fixed, and the outer recursion trains the model via reinforcement learning under the improved harness. Case studies span four scientific task families, covering researcher interaction, harness refinement, and model learning. The system is released as a research product; one report states it is publicly available at science-buddy.io. The work was listed on Hugging Face daily papers on 2026-09-14 and appeared on arXiv (cs.AI, cs.LG, cs.CL) on 2026-09-15. The two source reports are consistent, with no factual disagreements.

• System name: ScienceBuddy, an interactive scientific research workspace released as a research product to the scientific community
• Core method: recursive-in-recursive self-improvement coupling harness evolution with model reinforcement learning
• Inner recursion: harness evolution with the model held fixed
• Outer recursion: model reinforcement learning under the improved harness
• Data pipeline: converts researcher requests, feedback, and execution evidence into training tasks and evaluation rubrics
• Evaluation: case studies span four scientific task families, covering researcher interaction, harness refinement, and model learning
• Availability: publicly released at science-buddy.io (stated in the Hugging Face report)
• Coverage dates: Hugging Face daily papers 2026-09-14; arXiv cs.AI/cs.LG/cs.CL 2026-09-15