The Reinforcement Learning with Axiomatic World Modeling (RLAWM) method penalizes divergence between a model's internal physics model and true physical laws to prevent 'modeling drift' in reasoning.
HOW THIS AFFECTS YOU
●
researcherThis approach addresses the tendency of models to fabricate physical laws when reasoning in axiomatic domains.