Lunch talk on Sep. 14, 2026

Executable Scientific Verification with AI Agents: Formalizing Expert Reasoning in Astronomical Spectroscopy

Speaker: Haosong Wang (SHAO)

Venue: SWIFAR Building 2111

Time: 12:45 PM, Monday, Sep. 14, 2026

Abstract: 

Artificial intelligence is becoming increasingly capable of predicting, classifying, and interpreting astronomical observations. The emergence of AI agents extends these capabilities beyond individual prediction tasks: an agent can combine models, scientific software, domain knowledge, and iterative reasoning to carry out multi-step scientific workflows. This raises a broader question for astronomy: how can such systems participate in scientific analysis while keeping their conclusions transparent, testable, and scientifically reliable? In this talk, I will first give a brief introduction to AI agents and discuss how they differ from conventional machine-learning models and standalone large language models. I will then introduce our ongoing work on FORMA (Formalized Observational Reasoning with Auditable Decisions), an agent-based framework that formalizes expert reasoning as an executable astronomical spectrum verification process. Using the DESI visual inspection catalogue as a testbed, FORMA extracts observational evidence, constructs and tests alternative hypotheses, applies physical consistency checks, and records the reasoning process leading to each decision.