Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered

Aryo Pradipta Gema; Neel Rajani; Rohit Saxena; Wai-Chung Kwan; Pasquale Minervini

title

authors

Aryo Pradipta Gema, Neel Rajani, Rohit Saxena, Wai-Chung Kwan, Pasquale Minervini

date

2026-08-29

venue

arXiv (preprint, not peer reviewed)

arXiv:2608.29464

affiliation

EdinburghNLP - Natural Language Processing Group at the University of Edinburgh

upvotes

13 (Hugging Face Daily Papers, 2026-11-01)

links

abs pdf hf code

abstract

Chain-of-thought (CoT) monitoring assumes that reasoning traces faithfully record the information that shapes a model's answer. Existing faithfulness tests often place explicit bias cues in the user message, while agents may encounter preferences through tool returns or raw artifacts. We introduce FACE-Eval (Faithful Attribution of Cue Effects Evaluation), a 5,100-sample evaluation that varies cue location (user message or tool return) and explicitness (direct summary or raw artifact). We measure verbalized commitment among cue-following answers and unverbalized adoption among all cued samples. We evaluate 15 open-weight models from eight families, with total parameters ranging from 4B to 1.60T. Every model has lower verbalized commitment for tool-return than user-message cues and for implicit than explicit cues. Unverbalized adoption is higher for tool-return cues on all 15 models and for implicit cues in 28 of 30 model-channel comparisons. A source-attribution prompt narrows the channel gap on seven models, sometimes by increasing user-channel unverbalized adoption, while telling models that their reasoning will be monitored does not reliably close the gap. We also use two transcript monitors (GPT-5.6-Luna and GPT-4o-mini) to detect preference adoption in the largest model of each family. Across 32 model-channel-explicitness cells, higher unverbalized adoption is associated with lower detection ability for both monitors (Pearson r=-0.54 and r=-0.78, respectively). These results suggest that CoT monitoring may be less reliable when preference information arrives through tools or must be inferred from raw artifacts, within the single-call, prefilled-tool setting tested here.

@misc{gema2026chain,
  title={Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered},
  author={Aryo Pradipta Gema and Neel Rajani and Rohit Saxena and Wai-Chung Kwan and Pasquale Minervini},
  year={2026},
  eprint={2608.29464},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2608.29464}
}