Скачать или смотреть Physics Steering: Controlling Simulation Models

In this AI Research Roundup episode, Alex discusses the paper: 'Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model(2511.20798v1)' This paper explores whether large physics-focused foundation models learn human-understandable internal concepts, similar to what mechanistic interpretability has revealed in large language models. The authors extract activation vectors from simulations in different physical regimes and compute delta representations that act as directions for specific physical concepts. By injecting these concept directions back into the model during inference, they can causally steer its predictions, adding or removing particular physical features from simulations. The results suggest that these models capture generalized physical principles rather than relying only on superficial patterns. Paper URL: https://arxiv.org/pdf/2511.20798 #AI #MachineLearning #DeepLearning #PhysicsModels #MechanisticInterpretability #CausalControl #FoundationModels