ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
ScienceBuddy is introduced and released as an interactive scientific research workspace. It supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its core is recursive-in-recursive self-improvement, coupling harness evolution with model reinforcement learning. The inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Case studies cover researcher interaction, harness refinement, and model learning, with benchmark cases spanning four scientific task families. ScienceBuddy is released as a research product.
ScienceBuddy is an interactive scientific research workspace that brings continually improving scientific agents into researchers' everyday workflows. It supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning: the inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation. Case studies of researcher interaction, harness refinement, and model learning are presented, with benchmark cases spanning four scientific task families. By releasing ScienceBuddy as a research product, the paradigm is made available to the scientific community, taking a step toward discovery intelligence: scientific AI that advances through sustained collaboration with researchers and evolves alongside the research it supports.
The recursive-in-recursive self-improvement paradigm couples harness evolution with model reinforcement learning. The inner recursion improves the harness with the model fixed, while the outer recursion trains the model under the improved harness. Harness evolution shapes training experience, and model learning creates new opportunities for harness adaptation. This approach enables continual learning from researcher requests, feedback, and execution evidence, transforming them into tasks and evaluation rubrics.
ScienceBuddy is released as a research product, making the recursive-in-recursive self-improvement paradigm available to the scientific community. This represents a step toward discovery intelligence, where scientific AI advances through sustained collaboration with researchers and evolves alongside the research it supports. The release may influence how interactive scientific agents are developed and adopted in research workflows.
ScienceBuddy is released as a research product, potentially creating value by improving researcher productivity and enabling continual learning in scientific workflows. Its availability to the scientific community may lead to partnerships, further development, or commercialization opportunities in scientific AI tools.
Observable next signals include adoption of ScienceBuddy by researchers, publication of follow-up studies on recursive-in-recursive self-improvement, and expansion of benchmark cases beyond the four scientific task families. Further development may focus on scaling the paradigm to other domains and improving harness adaptation mechanisms.