← back
Autonomous Agents for Scientific Tasks - Sina Shahandeh, Radicait
Takeaway
Explicit problem decomposition can help autonomous research agents generate ideas beyond incremental code optimization.
Summary
- Scientific coding agents can optimize implementations but plateau when they exhaust useful research hypotheses.
- Radicait explores translating CT scans into synthetic PET images using an encoder-decoder model and image-fidelity metrics.
- A 2.5D model that stacks CT slices as channels illustrates how local hyperparameter searches can miss larger architectural changes such as 3D convolutions.
- Generate linked hierarchical documents covering data, architecture, losses, metrics, and code, then explicitly ask the agent to propose hypotheses across that hierarchy.
scientific-agentshypothesis-generationmedical-imaging
Original description
There has been much work on Autoresearch where the objectives are coding puzzles, toy optimization problems, or static supervised-learning ML tasks. However, for an autonomous agent to assist with a scientific discovery task, the problems must come from real measurement data of the world and they are highly open-ended, requiring a scientific method in the solution loop. In this talk, we show scenarios where the agent has to search over methods, priors, data preprocessing, model classes, and hyperparameters while learning from intermediate failures. Often, a step change in agent performance comes from forming an appropriate scientific hypothesis about how the physical system behaves, implementing that hypothesis correctly in a mathematical model, and executing it on the existing real data. We show how an ontology-based memory system is used in the harness to assist with the hypothesis generation that is key to the agent's success. All demonstrations come from real scientific problems solved in industrial and applied research settings. Speakers: Sina Shahandeh (Radicait): Sina is a cofounder/CTO at Radicait. His background is in scientific computing (PhD) and has lead data/AI function in 3 scale ups (with $770M exit) as well as founding 3 companies. X/Twitter: https://x.com/SinaShahandeh LinkedIn: / sinashahandeh