Circuit Insights: Towards Interpretability Beyond Activations

Researchers introduce WeightLens and CircuitLens, two new methods for analyzing neural network interpretability that go beyond traditional activation-based approaches. These tools aim to provide more systematic and scalable analysis of neural network circuits by interpreting features directly from weights and capturing feature interactions.

• →WeightLens interprets neural network features directly from learned weights without requiring external explainer models or datasets.
• →CircuitLens captures how feature activations arise from component interactions, revealing circuit-level dynamics missed by activation-only methods.
• →The methods address limitations of existing automated interpretability approaches that rely heavily on external LLMs and dataset quality.
• →Both tools are designed to increase interpretability robustness while maintaining efficiency for scalable mechanistic analysis.
• →The research builds on transcoders that separate feature attributions into input-dependent and input-invariant components.