Physicists have a dark matter problem: too many viable theories, not enough ways to tell them apart experimentally. A new AI framework called hAIthem aims to narrow that gap.
Researchers built hAIthem by pairing a reinforcement-learning agent with large language models. The RL agent, built as an autoregressive transformer called a Large Lagrangian Model, is pretrained on roughly 1 billion tokens drawn from about 10,000 Lagrangians - the mathematical descriptions physicists use to specify a theory. It then plays what the team calls a Battleship-style game against established physics simulation tools, probing which regions of a theory's parameter space survive current experimental constraints. A decision tree sorts the surviving regions, and when RL alone can't separate similar-looking cases, LLM agents compete to propose distinguishing observables.
This matters because dark matter research is bottlenecked less by theory and more by what's actually measurable. Gravitational evidence confirms dark matter exists but says little about which of the many candidate models is correct. An automated system that can comb through high-dimensional parameter spaces and flag overlooked observables could point experimentalists toward measurements worth funding, rather than leaving that work to manual trial and error.
In early tests on a simplified model with a single dark scalar particle, hAIthem's RL search beat a standard evolutionary-algorithm baseline, turning up a wider and more physically diverse set of viable regions - including a proposed use of existing kinematic techniques in paleo-detectors. It's a proof of concept, not a discovery, and the real test will be whether any of these proposed observables survive contact with an actual experiment.