Gen-Verse/ScienceBuddy

What it solves

ScienceBuddy is an interactive scientific agent designed to help researchers analyze scientific papers, data, and figures. It addresses the challenge of conducting complex scientific discovery by providing a workspace where an agent can retrieve literature, inspect biological sequences, and organize evidence into tables, while allowing the researcher to refine the analysis through a continuous dialogue.

How it works

The project consists of two primary components: a researcher-facing workspace and a research framework for agent improvement.

1. Scientific Workspace: The agent uses a suite of 224 tools across 22 functional modules (specializing in biomedicine, genomics, and pharmacology) to execute tasks. Researchers can upload documents or images, ask questions, and track the agent's execution trajectory and tool outputs.
2. Double-Recursive Self-Improvement (RSI): This is a learning framework that iteratively improves both the agent's guiding harness (the Python programs that direct the agent) and the task model itself. It uses a cycle where the harness is refined based on interaction feedback, and the model is then trained using SkyRL GRPO to perform better under that updated harness.

Who it’s for

It is primarily for scientific researchers, particularly those in the biomedical and life sciences, who need to automate the evidence retrieval and analysis of complex scientific materials.

Highlights

• Multimodal Input: Supports papers, tables, biological sequences, and scientific images.
• Extensive Toolset: Includes 224 tools for genomics, cancer biology, bioimaging, and database queries.
• Harness-Model Co-evolution: A recursive framework that improves both the operational procedures (harness) and the underlying LLM (model) through alternating learning stages.
• Transparency: Provides a "Trajectory" view to inspect the exact tool inputs and outputs used to reach an answer.