05/04/2025
Would two AI scientists disagree with each other, even if trained on the same data?
After seeing classical physics, AI scientists disagree at first but converge to known theories (Lagrangian/Hamiltonian) when data become diverse.
v/ Lu
Do Two AI Scientists Agree?
Xinghong Fu, Ziming Liu, Max Tegmark
In the world of science, theories emerge, evolve, and sometimes fade as new experimental data shapes our understanding. But what about AI scientists—artificial intelligence models trained to discover scientific principles? Do they arrive at the same conclusions when given the same task, or do they develop their own distinct theories?
This study explores that question by training AI models on physics problems to see whether they independently arrive at the same scientific laws. Just as human scientists once debated competing theories before settling on the best explanation, these AI models sometimes converge on a single theory—but at other times, they split into distinct groups, each favoring a different perspective.
To investigate this, researchers developed an approach called MASS, using Hamiltonian-Lagrangian neural networks as AI scientists. By training these AI models on standard physics problems and analyzing results across multiple training runs (simulating different "scientific communities"), they observed a fascinating shift: AI models tend to favor Hamiltonian mechanics in simpler cases but switch to Lagrangian mechanics when dealing with more complex systems.
Moreover, the process is highly sensitive to initial conditions, meaning that slight variations in training setups can determine which theory gains prominence—mirroring how human scientific paradigms rise and fall over time. Beyond improving our understanding of AI learning, this approach could also enhance interpretability in high-dimensional scientific problems, pushing the boundaries of AI-driven discovery.
