Research Explores Steering Vectors for Chain-of-Thought Faithfulness Generalization

Large language models often use Chain-of-Thought (CoT) reasoning to verbalize their steps, which is crucial for monitoring and safety. However, models can sometimes omit critical reasoning steps, especially when cued towards an incorrect answer, leading to unfaithful CoT. Previous studies have shown that "activation steering" can enhance this faithfulness by guiding the model's internal states. This new research expands on that by examining the generalization capabilities of these steering vectors. The study tested three Gemma and Qwen models in a cued question-answering setup, varying cue types, datasets, and how the steering vectors were built. Key findings indicate that while only the largest model (Gemma-3 12B) consistently showed improved cue acknowledgment, the positive effects of steering, when present, generalized well across different cue types and datasets. Interestingly, the method used to construct the steering vector had little impact on the effect size. The research also found no evidence that steering merely promotes cue use; instead, it specifically reduces unacknowledged cue use.