Researchers Introduce Capsule Lens to Track Concept Geometry in AI Representations

Researchers have proposed Capsule Lens, a new framework designed to advance mechanistic interpretability by directly characterizing how concepts occupy representation spaces in machine learning models. The framework models the region a concept occupies using a capsule geometry with interpretable parameters fitted in closed form.

The authors applied Capsule Lens across both static and dynamic model representations. The evaluation included three case studies tracking representation drifts across different training regimes: CLIP pretraining, reinforcement learning post-training on visual question answering, and reinforcement learning post-training on mathematical reasoning.

The findings revealed that CLIP pretraining induces broad network-wide structural changes, whereas reinforcement learning post-training results in localized, concept-specific modifications.