ML-Astrobiology Initiative Virtual Seminar: Human-Machine Collaboration for Scientific Hypothesis Generation: Lessons from AstroAgents
Hosted by the ML-Astrobiology Initiative at NASA Ames Research Center
The speaker for this event will be Dr. Amirali Aghazadeh, from the Georgia Institute of Technology.
ASTROAGENTS
Human-Machine Collaboration for Scientific Hypothesis Generation: Lessons from AstroAgents
Large language model (LLM) agents can now read scientific literature, analyze data, and propose candidate hypotheses, raising a practical question: which parts of scientific reasoning should be delegated to machines, and which should remain with the scientist? For in-situ exploration, greater autonomy may be necessary because of communication delays and limited bandwidth; on Earth, however, this is a choice.
I will discuss this question through AstroAgents, a multi-agent system that generates hypotheses from untargeted GC×GC mass spectrometry data on carbonaceous chondrites and terrestrial Mars-analogue samples, building on our LifeTracer framework for distinguishing abiotic from biotic organics.
I will describe how the system moves from data analysis and planning to hypothesis generation, literature review, and critique, and what expert evaluation of more than 100 generated hypotheses revealed about novelty, plausibility, and common failure modes.
I will then place these results in the context of recent co-scientist agentic systems and discuss what they suggest about the roles of data, evaluation, and human oversight. I will close with practical lessons for using these systems in scientific research and for deciding when greater autonomy is appropriate.
Questions/Comments Contact Dr. Ryan Felton ryan.c.felton@nasa.gov