Beyond a Single Direction: Chain-of-Thought Disrupts Simple Steering of Refusal
The paper discusses how chain-of-thought (CoT) in large reasoning models (LRMs) complicates the control of refusal mechanisms. It highlights that refusal is influenced by both the CoT and residual stream activations, making LRMs more resilient to simple activation-level interventions. The findings suggest that the CoT can independently carry compliance signals, indicating a need for more nuanced control strategies.
- ▪Refusal in LRMs is influenced by chain-of-thought (CoT) and residual stream activations.
- ▪Activation steering reverses refusal in only 39% of cases when CoT is fixed, but increases to 70% when CoT is removed.
- ▪A two-stage intervention can reverse refusal in 94% of cases, demonstrating the CoT's role in compliance.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.26772 (cs) [Submitted on 26 May 2026] Title:Beyond a Single Direction: Chain-of-Thought Disrupts Simple Steering of Refusal Authors:Kia-Jüng Yang, Dominik Meier, Jiachen Zhao, Terry Ruas, Bela Gipp View a PDF of the paper titled Beyond a Single Direction: Chain-of-Thought Disrupts Simple Steering of Refusal, by Kia-J\"ung Yang and 4 other authors View PDF HTML (experimental) Abstract:Large reasoning models (LRMs) generate chain-of-thought (CoT) traces before producing final outputs, introducing a dynamic internal state that may complicate control mechanisms such as refusal. Unlike instruction-tuned LLMs, where refusal is mediated by a single directional subspace, refusal in large reasoning models (LRMs) additionally depends on the CoT.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
