Interpreting and Steering State-Space Models via Activation Subspace Bottlenecks

TL;DR

This paper identifies activation subspace bottlenecks in state-space models, introduces a simple steering intervention to improve performance, and validates the bottlenecks' impact through architecture modifications, advancing interpretability and steerability.

Contribution

It uncovers activation bottlenecks in SSMs, proposes a scalar multiplication intervention, and demonstrates performance gains and interpretability improvements.

Findings

Intervention improves performance by 8.27% on average across benchmarks.

Identified bottlenecks hinder model performance, validated by architecture modifications.

Stable-Mamba architecture achieves long-context performance gains when retrained.

Abstract

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 7 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.