I had two papers accepted by the ICML 2026 AI for Science Workshop!
They focus on the gap between what AI agents plan in silico and what they can actually do in physical labs. There is a big discussion about autonomous AI scientists and their danger, but physical execution is where things break, it’s harder to debug things in the physical world compared to code and math.
AI Scientist Agents and Biosecurity
First paper is AI Scientist Agents and Biosecurity: Capabilities, Risks, and Governance for Autonomous Labs.
Main arguments:
- AI scientist agents excel at digital tasks like planning, tool use, and protocol writing. However, they struggle with tacit wet-lab dexterity, perceptual disambiguation of inventories, and physical error recovery.
- Because physical bottlenecks remain high, a lone actor working in an improvised setting gets little marginal help from these agents. Instead, the risk shifts to groups that can pair agentic planning with structured execution environments, such as automated workcells or remote cloud labs.
- Effective governance should focus less on restricting model access and more on physical bottlenecks, such as improving synthesis screening, auditing remote-access automated facilities, and investing in AI-driven defensive response.
A Dataset for Scientific Debugging
Second paper is Experimental Attempts in Electronic Lab Notebooks: A Dataset Proposal for Scientific Debugging.
Scientific publishing suffers from a clear publication bias where we generally only share successful protocols. Agents need to learn how experiments fail and what to do to debug them. I propose a dataset schema for run-level experimental attempt records. This schema documents the initial goal and hypothesis, the executed run text, the status (success, failure, or partial), and the troubleshooting response.
I have already assembled pilot corpora from synthetic biology wiki records on iGEM and other public sources. The next step is building extraction pipelines to map native fields from existing Electronic Lab Notebooks (ELNs) into this schema, without forcing researchers to change their local workflows.