ScienceBuddy: Recursive-in-Recursive Self-Improveme... | AI Research

ScienceBuddy is an interactive research workspace designed to help scientific agents learn and improve through ongoing collaboration with researchers. Instead of treating scientific tasks as isolated, one-off requests, ScienceBuddy captures the dialogue, feedback, and execution evidence from a researcher’s daily work to create a cycle of continuous improvement. By integrating a flexible "harness"—a set of instructions and tools that guide the agent—with a learning model, the system evolves alongside the research it supports.

Recursive-in-Recursive Self-Improvement

The core innovation of ScienceBuddy is a "recursive-in-recursive" paradigm that links two distinct learning processes. The "inner" recursion focuses on the agent's harness: while the model itself remains fixed, an auxiliary system analyzes past performance to refine the agent’s instructions, skills, and context settings. The "outer" recursion then takes this improved harness and trains the model using reinforcement learning. Because these two processes are coupled, the harness shapes the training experience, and the updated model creates new opportunities for the harness to be adapted further. The same ai systems question is explored in A Unified Physics-Aware Quantum Machine Learning..., which adds a research perspective.

Turning Collaboration into Tasks

ScienceBuddy transforms the natural back-and-forth between a researcher and an AI into structured learning data. When a researcher asks a question, inspects a result, or requests a revision, the system uses these interactions to define specific scientific tasks and evaluation rubrics. These rubrics act as a scorecard, allowing the agent to measure its own success against objective criteria rather than just relying on the researcher’s approval. These "Harbor tasks" provide a consistent way to test the agent’s performance as it continues to learn.

A Multimodal Scientific Workspace

To support complex research, ScienceBuddy provides a unified environment that combines 224 specialized tools across fields like genomics, pharmacology, and bioimaging. The workspace is designed to be persistent, meaning it retains documents, images, tables, and biological sequences across different sessions. By separating the agent’s "harness" from the underlying scientific infrastructure, the system allows researchers to see exactly how the agent is being instructed to solve problems, making the agent’s reasoning process transparent and easier to refine. The same ai evaluation question is explored in CUA-Universe, which adds a research perspective.

Evolution Through Sustained Interaction

The system is designed to operate in the background, allowing the agent to improve while the online service remains available to the researcher. As the agent gains experience, the system automatically calibrates the difficulty of new tasks to ensure the agent is always being challenged appropriately. By retaining all previous versions of its harness and environment, ScienceBuddy ensures that the agent’s capabilities grow steadily, turning the research workspace into a collaborative environment where the AI evolves as a partner in the scientific discovery process. The same ai evaluation question is explored in Procedural Graphs, which adds a research perspective. as detailed in the full paper on Arxiv