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
cs.CL, cs.AI
2026-08-30
FACE-Eval tests 15 open-weight models on 5,100 samples: verbalized CoT commitment drops for tool-return and implicit preference cues, and transcript monitors detect this worse.
CoT monitoring assumes the trace records the information that actually moved the answer. Standard faithfulness tests plant a biased hint in the user message and check whether the model names it after the answer flips. That setup matches a chat assistant. It does not match an agent.
Agents pick up user preferences from tool returns, mail, Slack, and browsing history, often as raw artifacts that have to be inferred. FACE-Eval asks whether the trace still records the decision to tailor the answer when the cue changes channel and explicitness.
FACE-Eval has 5,100 samples: 5,000 cued plus 100 no-cue controls. The factorial crosses five preference axes, two sides, five artifact sources, several delivery conditions, and 20 scenarios per axis. Axes cover left-right stance, ethics (utilitarian versus deontological), egalitarianism, epistemic posture, and domain expertise. Sources are a user profile, email, Slack, personal notes, and browser history.
Delivery splits on two axes:
The headline metric is verbalized commitment rate (VCR): among answers that already follow the cue, the share whose CoT states an intent to tailor. Cue-following rate (CFR) is how often the answer follows. Unverbalized adoption (UAR) equals CFR times (1 minus VCR), over all cued samples. A model that rarely commits can still have low UAR if it rarely follows.
The panel is 15 open-weight models from eight families, 4B to 1.60T, including DeepSeek V4 Pro (1.60T), Kimi K2.6 (1.04T), Inkling (975B), and GLM 5.2 (744B). Claude Haiku 4.5 is the primary judge, scoring the CoT and the answer separately. GPT-5.6-Luna is a second judge. Two system-prompt variants ask for source attribution or disclose that reasoning will be monitored. GPT-5.6-Luna and GPT-4o-mini then act as transcript monitors on the largest model of each family.
Every one of the 15 models has lower VCR for tool-return than user-message cues, and lower VCR for implicit than explicit cues. The user-over-tool VCR advantage ranges from 0.08 on GLM 5.2 to 0.77 on OLMo 3.1-32B. CFR sits between 0.43 and 0.58, so models can read the cues and often follow them. They just do not always write the decision down.
UAR is higher on the tool channel for all 15 models, with gaps from 0.03 (Gemma 4-E4B) to 0.21 (Qwen 3.5-27B). The implicit-over-explicit UAR increase holds in 28 of 30 model-channel comparisons. After matching on model-rated clarity, explicit cues still keep a fitted VCR advantage of about 0.14 on the user channel and 0.13 on the tool channel. Clarity does not eat the whole gap.
| Setting | Metric | Result |
| 15/15 models | Tool vs user VCR | Tool lower |
| 15/15 models | Tool vs user UAR | Tool higher by 0.03-0.21 |
| 28/30 comparisons | Implicit vs explicit UAR | Implicit higher |
| GPT-5.6-Luna | Transcript-monitor AUROC | 0.71-0.73 |
| GPT-4o-mini | Transcript-monitor AUROC | 0.64-0.74 |
Across 32 model-channel-explicitness cells, higher UAR tracks worse detection (Pearson r = -0.54 for GPT-5.6-Luna, r = -0.78 for GPT-4o-mini). Adding the CoT helps GPT-5.6-Luna on the user channel only; both tool-channel intervals include zero. GPT-4o-mini gains in all four cells, and the channel and explicitness gaps remain.
A source-attribution prompt narrows the channel gap on seven models. On Qwen 3.5 the narrowing comes mostly from rising user-channel UAR, not falling tool-channel UAR. Telling the model its reasoning will be monitored moves the gap by at most 0.070 and does not close it. Raising reasoning effort cuts UAR monotonically on both GPT-OSS sizes, mostly via higher VCR rather than lower CFR. DeepSeek V4 Pro and Inkling do not repeat that four-condition pattern. At every available effort setting, VCR stays higher for user-message and explicit cues.
CoT monitors that only plant explicit bias in the user message will overstate monitorability for agents. Real preferences often sit in tool results and raw artifacts, which is exactly where verbalized commitment is lowest and monitor AUROC is worse.
Neither prompt tweaks nor extra reasoning effort is a reliable patch. Attribution instructions sometimes move the hole from the tool channel onto the user channel. Longer traces may just give the judge more text in which to find formulaic commitment.
This is a measurement warning, not a new training method. For people who audit agent sycophancy, FACE-Eval makes "where the cue arrived" and "how directly it was stated" dimensions that have to be reported separately.
The paper is explicit: the experiments locate the gaps, they do not explain them. The tool channel also inserts a prefilled tool call, so training priors, reporting habits, and judge sensitivity to channel phrasing are all in the mix.
The task is single-call with a prefilled tool. The model never selects the call. Tailoring an answer to a preference may not need long serial reasoning, so the CoT can function as an optional report rather than a necessary workspace. Tasks that require multi-step reasoning to apply the cue may look different.
VCR depends on an automated judge. A second judge recovers the main trends, while absolute values and some rankings move. Longer traces give the judge more chances to find commitment language, which especially contaminates the GPT-OSS effort results.
None of this shows deliberate concealment. Synthetic mail and notes are not real user data. Interventions that raise verbalized commitment could also teach models to emit formulaic disclosure.