Scientific Reasoning Agents

Project Summary

Developing agentic AI systems that combine large language models (LLMs) with scientific domain tools and data to plan, simulate, and iteratively propose hypothesese and design experiments. Building on recent advances in reasoning agents and autonomous labs, we will create end-to-end “scientific agents” that (i) generate testable hypotheses, (ii) call specialist software (e.g., simulation, cheminformatics, statistical design packages), (iii) propose and schedule experiments, and (iv) learn from results via active learning/Bayesian optimisation loops. We will validate agents on open, auditable in-silico benchmarks and, where possible, with closed-loop experiments (in-silico and hardware-in-the-loop) using our autonomous labs.

Potential Supervisors

• Dr Liam Atkinson (Research Engineer, EIT)
• Dr Ira Ktena (Research Scientist, EIT)
• Dr Ben Chamberlain (Research Scientist, EIT)
• Dr Danilo Jimenez Rezende (Research Scientist, EIT)
• Additional Supervisor(s) from the University of Oxford

University DPhil Courses 

• DPhil in Chemistry  
• DPhil in Statistics  
• DPhil in Mathematics  
• DPhil in Computer Science  
• DPhil in Engineering Science

Relevant Background Reading

Reasoning + Acting with Tools

• Yao et al., ReAct: interleaving thoughts and tool calls for reasoning and information-seeking
• Schick et al., Toolformer: self-training LMs to decide when/what/how to call APIs
• Yao et al., Tree of Thoughts: deliberate multi-path search during inference

LLM Agents for Science

• Bran et al., ChemCrow
• Boiko et al., Coscientist (Nature)

Closed-LoopExperimentation

• Bendersky et al., Closed-loop materials discovery (CAMEO):on-the-fly Bayesian active learning at a synchrotron.

Evaluation & Reliability

• AgentBench and related real-computer benchmarks

‍