Human-Centred Agentic Science explores how AI agents can support researchers across the stages of scientific discovery, from literature research and hypothesis generation to implementation, evaluation and writing. The aim is to let AI take on time-consuming work while researchers guide the process, apply their expertise and make the important decisions — not to automate science end-to-end. Read more about Human-Centred Agentic Science ↗
Below are more specific project ideas, all under the Human-Centred Agentic Science:
Idea 1: Human-Centred Agentic Science for Literature Research
Develop tools that help researchers explore a field, understand relationships across papers and identify research gaps. Agents can iteratively suggest topics and relevant work, while researchers steer the exploration with feedback and draw their own conclusions. This could build on focused literature collections and systems such as vitaLITy 2 and AwesomeLit. Read more about this project idea ↗
Idea 2: Human-Centred Agentic Science for Quantitative Research
There are many different types of quantitative researches. Here we use human-guided machine learning for automated trading as an example: combining market data with news sentiment, exploring reinforcement-learning and time-series anomaly-detection approaches, and building interactive tools to compare and tune models. The focus is on helping researchers inspect performance and guide experimentation, rather than assuming a model will reliably make profitable trades. Read more about this project idea ↗
Idea 3: Human-Centred Agentic Science for Qualitative Research
Explore how LLMs can assist with repetitive parts of qualitative analysis, such as first-pass coding and checking consistency across themes or coders. Researchers review the evidence and recommendations, retaining control over interpretation and decisions. One example is Co-Refine, which flags possible coding inconsistencies during analysis. Read more about this project idea ↗