Beyond a Single Direction: Chain-of-Thought Disrupts Simple Steering of Refusal
Factual evidence
What the source reports
Research finds Chain-of-Thought (CoT) reasoning in LLMs complicates refusal control, making activation steering less effective than with non-CoT models.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.LG — Machine Learning · RESEARCH
- Published by source
- 21 October 2026
- Collected by OneBench
- 21 Jul 2026, 03:06 UK
Stored source excerpt
arXiv:2605.26772v1 Announce Type: cross Abstract: Large reasoning models (LRMs) generate chain-of-thought (CoT) traces before producing final outputs, introducing a dynamic internal state that may complicate…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
The internal complexity of CoT reasoning makes controlling LLM behavior, especially refusal, more difficult and less predictable for enterprise safety and compliance requirements.
Do what
Your model risk and safety teams must account for increased complexity in steering and ensuring refusal behaviors for models using Chain-of-Thought, impacting validation and alignment efforts.