mechanisticinterpretability

Realtime unit labeling and debugging demo: streaming NeuronCards with confidence, stable identity under drift, and active probing for AI interpretability.

A research library for mechanistic interpretability and Theory of Mind in large language models

Investigating implicit reasoning in GPT-2-like Transformers through mechanistic interpretability (Logit Lens and Causal Tracing). Building on NeurIPS 2024 research, this work demonstrates how a targeted training-data intervention overcomes the OOD compositionality gap in a controlled two-hop reasoning task.