Vibe Coding DiscoverScienceBuddy
View on GitHubScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
Research code and preview for ScienceBuddy, an interactive scientific agent workspace plus a double-recursive self-improvement experiment that alternates harness refinement with SkyRL GRPO model training on frozen scientific tasks.
Use Cases
Interpret scientific figures and organize related evidenceRetrieve literature and protein/database records for a research questionBuild evidence tables that separate retrieved records from gapsRun biomedical analyses (genomics, pharmacology, bioimaging) through 224 toolsInspect agent trajectories, tool inputs/outputs and artifactsImprove a Python agent harness via recursive self-improvementTrain a task model with GRPO driven by verifier scoresReproduce a 715/90/90 train/val/test scientific agent benchmark
Built With
- Language
- Python
- Frameworks
- SkyRL · vLLM · PyTorch · FlashAttention · pytest · uv · jsonschema
Tags
scientific-agents · interactive-agent · recursive-self-improvement · agent-harness · llm-agents · reinforcement-learning · grpo · agent-evaluation · biomedicine · tool-use · research-automation · trajectory-inspection · verifier · python