CRL uses reinforcement learning to select sparse autoencoder (SAE) features for steering language models at each token, revealing which features impact outputs. It includes adaptive masking for diverse features and enables analysis like branch point tracking and layer-wise comparisons. Tested on Gemma-2 2B, it improves benchmarks while providing interpretable logs.
Key Points
- 1.RL policy selects SAE features per token
- 2.Tracks branch points and critic trajectories
- 3.Syntactic features early, semantic later layers
Impact Analysis
Advances mechanistic interpretability by combining static analysis with dynamic interventions. Enables precise model steering and error diagnosis. Complements existing SAE methods for better AI understanding.
Technical Details
Trains policy on Gemma-2 2B with MMLU, BBQ, GSM8K. Uses adaptive feature masking for interpretability. Reveals layer-specific feature types.