Department of Chemistry - PhD Student Seminar - LLM-Based Multi-Agent Systems for Autonomous Reaction Discovery and Optimization

Student: Ms. Yuqi FANG

Department: Department of Chemistry, HKUST

Supervisor: Prof. Yong HUANG

Co-supervisor: Prof. Jiean CHEN

Abstract

The exploration of chemical frontiers is bound by a multifaceted complexity: an astronomically large molecular space, a combinatorially explosive landscape of multi-dimensional reaction conditions, and a deluge of unstructured literature far exceeding human absorption capacity. While large language models (LLMs) promised to bridge this chasm through semantic proficiency, their autoregressive nature conceals an intrinsic vulnerability — a model trained to predict token probabilities is not natively bound to respect chemical topology, thermodynamics, or conservation laws, and this separation between textual fluency and physical reality inevitably surfaces as miscalibrated, confident-sounding hallucinations that fail the rigorous demands of hard science.

This seminar traces the paradigm shift from two parallel trajectories — tool augmentation for digital reasoning and automated laboratories for physical execution — to their current convergence into LLM-based Multi-Agent Systems (MAS), where specialized agents plan, act, observe, and reflect in a closed loop. Crucially, this convergence reveals a counterintuitive pattern: progress is determined not by model size or agent count, but by how the division of labor is structured. Ultimately, this structural optimization underpins a "Human-in-the-Loop" paradigm, where multi-agent architectures scale execution while human intuition arbitrates discovery.