Generalization of Steering Vectors for Chain-of-Thought Faithfulness

Chain-of-thought monitoring relies on the chain of thought being faithful, meaning that it reports the cues that determine a model's answer. We study whether steering vectors that increase cue acknowledgment generalize across cue types, datasets, and methods of constructing the vector, evaluated across three models (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B) in a cued question-answering setting. Cue acknowledgment is linearly decodable, with probe AUC that is high on MMLU and close to chance on GPQA. Steering increases acknowledgment, but the size of the improvement depends on the model and the dataset: the change is distinguishable from zero only for the largest model (Gemma-3 12B), where it is largest on MMLU (about $+0.26$), intermediate on BBH (about $+0.10$), and smallest on GPQA (about $+0.06$); the two smaller models show little change on any dataset. The improvement does not clearly depend on how the vector is constructed or how strongly it is applied: four construction methods, including one that uses no cued data, give comparable changes, and within each dataset the change is about constant across the steering strengths we tried. We do not have an explanation for the dependence on model and dataset. Separating acknowledgment from cue use shows that the increase appears whether or not the model selected the cued option, while the rate of cue use is roughly unchanged.