AI Inventor: Automating Scientific Discovery with LLM Agents

Large language models (LLMs) are everywhere, but how do we actually understand their capabilities and push them to their absolute limit? This talk walks through a concrete example of a state-of-the-art system for fully end-to-end autonomous AI science we are developing at the Jozef Stefan Institute. It ingests scientific literature, builds a knowledge graph, and detects blind spots where existing ideas have never been combined. LLM Agents then generate hypotheses, and an iterative discovery loop takes over: proposing conjectures, running real experiments, evaluating results, and refining until it arrives at a publishable contribution, complete with a full scientific paper and reproducible code. LLMs can be difficult to use at times and underwhelming, but with the right tools, verification, and prompting, they can create something previously thought impossible.

Adrian Mladenic Grobelnik began his involvement in computer science as a competitive programmer and has been doing AI research at the Jozef Stefan Institute since high school. He is currently completing his Master’s degree at the Jozef Stefan International Postgraduate School. He spent time at Stanford University working with Prof. Jure Leskovec on optimizing compound AI systems, completed a summer internship at CMU applying AI to healthcare, and collaborated with Stephen Wolfram on multi-agent LLM simulations. He has 10 scientific publications and a Best Paper Award; his work on optimizing compound AI systems was accepted at ICLR 2026.