Disentangling MLP Neuron Weights in Vocabulary Space
This work addresses the problem of scalable, fine-grained interpretation of language models for researchers and practitioners, offering a novel approach but with incremental improvements in interpretability methods.
The paper tackles the challenge of interpreting model weights in mechanistic interpretability by introducing ROTATE, a data-free method that disentangles MLP neurons in weight space to recover sparse, interpretable vocabulary channels, resulting in neuron descriptions that outperform activation-based baselines by 2-3x in comparisons.
Interpreting the information encoded in model weights remains a fundamental challenge in mechanistic interpretability. In this work, we introduce ROTATE (Rotation-Optimized Token Alignment in weighT spacE), a data-free method requiring no forward passes that disentangles MLP neurons directly in weight space. Our approach relies on a key statistical observation: neurons that encode coherent, monosemantic concepts exhibit high kurtosis when projected onto the model's vocabulary. By optimizing rotations of neuron weights to maximize their vocabulary-space kurtosis, our method recovers sparse, interpretable directions which we name vocabulary channels. Experiments on Llama-3.1-8B-Instruct and Gemma-2-2B-it demonstrate that ROTATE consistently recovers vocabulary channels that are faithful to the neuron's behavior. ablating individual channels selectively disables corresponding input activations or the promotion of specific concepts. Moreover, aggregating channel-level descriptions yields comprehensive neuron descriptions that outperform optimized activation-based baselines by 2-3x in head-to-head comparisons. By providing a data-free decomposition of neuron weights, ROTATE offers a scalable, fine-grained building block for interpreting LMs.