Interpreting and Steering State-Space Models via Activation Subspace Bottlenecks
This work provides a method for improving the performance of Mamba-family SSMs for researchers and practitioners working with these efficient language models, representing a strong specific gain in their application.
This paper addresses the interpretability and steerability of Mamba-family state-space models (SSMs) by identifying activation subspace bottlenecks. By simply multiplying the activations of these bottlenecks by a scalar during test-time, the authors achieved an average performance improvement of 8.27% across 5 SSMs and 6 benchmarks.
State-space models (SSMs) have emerged as an efficient strategy for building powerful language models, avoiding the quadratic complexity of computing attention in transformers. Despite their promise, the interpretability and steerability of modern SSMs remain relatively underexplored. We take a major step in this direction by identifying activation subspace bottlenecks in the Mamba family of SSM models using tools from mechanistic interpretability. We then introduce a test-time steering intervention that simply multiplies the activations of the identified bottlenecks by a scalar. Across 5 SSMs and 6 diverse benchmarks, this intervention improves performance by an average of 8.27%, without requiring any task-specific tuning. Finally, we validate that the identified bottlenecks are indeed hindering performance by modifying them to yield an architecture we call Stable-Mamba, which achieves long-context performance gains when retrained from scratch.