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.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2602.22719 [cs.LG] |
| (or arXiv:2602.22719v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2602.22719 arXiv-issued DOI via DataCite |
From: Vamshi Sunku Mohan [view email]
[v1]
Thu, 26 Feb 2026 07:46:42 UTC (7,614 KB)
[v2]
Thu, 21 May 2026 10:49:19 UTC (8,100 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。