acceptodds
Under review as a conference paper at ICLR 2027

Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings

Abstract

Chain-of-thought (CoT) monitoring is increasingly treated as an important safety layer for frontier reasoning models. Most monitorability evaluations study explicit-influence settings: setups where the prompt directly incentivizes the model to hide something, e.g., by instructing it to perform a hidden side-task. A complementary axis for CoT-monitor evaluations is implicit-influence settings, where the prompt contains no instruction to hide, but the model's behavior is still shaped by features of the task or context, e.g. an irrelevant detail about a candidate that biases a hiring rating. We introduce the first benchmark that directly compares CoT monitorability under the two regimes. We test how model behavior changes in the presence of a nudge to choose a particular options. The nudge is delivered either as a casual aside (implicit), or as a direct instruction to act on the nudge and to conceal having done so (explicit). The benchmark spans four task formats (binary choice, numeric rating, multiple-choice QA, open-ended coding) and seven frontier extended-thinking models. Under explicit influence, a CoT monitor detects 60–94% of behavior shifts: even models instructed to conceal it leak the instruction into their CoT. Under implicit influence, the same factors still shift behavior, but detection falls by 41–46 percentage points in two of our four settings. Realistic system-prompt additions (of the kind a developer might deploy to reduce off-topic bias) lower implicit detection further, to as low as 5%, while preserving the behavioral influence itself. These results suggest that monitorability estimates obtained in explicit-influence settings may over-estimate monitorability, and that monitorability can be further decreased by well-intentioned deployment choices.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
chain-of-thought monitoring, chain-of-thought faithfulness, AI safety, reasoning models, large language models, monitorability, evaluation benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_hqMEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the reliability of COT monitoring under implicit versus explicit influence. It introduces a matched benchmark across four task formats (binary choice, numeric rating, multiple-choice QA, and open-ended coding) where the same behavioral nudge is delivered either implicitly as contextual information or explicitly as an instruction to act on and conceal it. Across seven reasoning models, the authors find that implicit influences can remain behaviorally effective while being substantially less detectable in CoT than explicit ones, with detection further reduced by benign guidance system prompts. They also study load-bearing multi-hop variants, showing that requiring explicit reasoning can partially recover monitorability.

Critical strengths
  1. The paper is clearly written and easy to follow.

  2. The benchmark is relatively comprehensive. It covers four different task formats and evaluates seven frontier reasoning models from multiple providers.

Critical weaknesses

My main concern is the novelty of the work. Implicit influence is not a topic that has been absent from prior CoT monitoring research, as the paper itself acknowledges, and I am not fully convinced that introducing a matched explicit-vs.-implicit comparison is by itself a sufficiently novel contribution. The load-bearing reasoning experiments are interesting, but the paper also positions this phenomenon as connected to prior findings rather than fundamentally new. Correspondingly, many of the main observations (for example, that implicit influences are harder to detect than explicit instructions) feel fairly intuitive rather than especially surprising or informative. I do think the benchmark itself may be valuable, but I am not yet convinced by the novelty of the overall contribution.

I would encourage the authors to articulate more clearly what they see as the paper’s strongest novel contribution, or alternatively, what the most actionable takeaway is that was not already established by prior work.

I am currently leaning toward weak rejection, but I may adjust my score depending on the authors’ response regarding the novelty and contribution of the work.

Questions for authors
  1. Have you investigated why the gap between implicit and explicit influence varies so substantially across the four settings? To what extent do you think this variation is driven by the implicit/explicit distinction itself, versus properties of the specific task designs? Understanding which factor is more important would affect how broadly the main conclusion should be interpreted.

  2. How sensitive are the results to prompt formulation? In particular, for the same underlying nudge, do different phrasings lead to different behavioral-shift or monitor-detection rates, and have you observed any systematic patterns? Similarly, have you evaluated the robustness of the adversarial and guidance system-prompt results to alternative prompt formulations?

Suggestions:

  1. I suggest moving more details about the benchmark statistics and the construction of the nudges into the main paper, as I think this information is important for understanding and interpreting the benchmark.

  2. The authors may also want to discuss some recent work on CoT monitoring benchmarks, such as MonitorBench [1].

[1] Wang et al., “MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models,” COLM, 2026.

Reviewer_fyTiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

At its core, the paper is showing that CoT is an incomplete representation of the model's decision making process. Prior works have shown this by showing that LLMs are influenced by explicit cues present in the prompt and this influence is omitted from the CoT sometimes; however, modern LLMs seem to have been patched against this particular form of exhibiting CoT unfaithfulness. Thus, this develops some additional settings and uses those to demonstrate that LLMs are influenced when cues are present implicitly which they nevertheless omit from their CoTs.

Critical strengths

The paper is generally a high-effort on the part of the authors. The experiments are comprehensive and there is I believe lots of raw data that authors have likely collected that could be interesting. However, the paper is lacking in the analysis of this data due to which I am unable to rate this paper as highly as I would have liked to otherwise.

Critical weaknesses
  • The paper has two main findings: (a) LLMs get subtly influenced by cues present in the prompt, (b) CoTs is an incomplete description of model's decision making process. Neither of these are novel in my opinion. In general, the novelty seems limited relative to prior works (e.g., https://arxiv.org/abs/2503.08679).

  • I also have a general issue with the settings used in this paper where the oracle process is allowed to use intervention experiments to establish what factors influenced the model's cognition, however, model is expected to have sufficient level of introspection and meta-awareness of these factors and mention them in its CoT. I am just not sure that this is feasible given current models are not that good at introspection.

  • The work only uses a single monitor which is insufficient to establish the negative finding. In order to show that CoTs lack the relevant information, it would be good to show that monitorability failures occur across monitors and are consistent across monitors.

  • The writing is okay-ish; a motivated reader can generally understand things but there is lots of room for improvement. In general, I found it hard to read the paper and often had to go back or read sections twice to fully understand things. I think this is probably because of use of AI (likely Claude) models in writing who are unfortunately not best at delivering dense information in a simple way and needlessly convolute things. There is also over-representation of numerical results in the text of the paper, without adequate analysis or discussion (this is a very standard AI thing to do and makes for a poor reading experience for humans who often rely more on qualitative trends obvious from the plots than on precise numerical values). I would recommend authors to try and orient their paper around interesting findings / analysis and less as a barrage of information. Additionally, things are too distributed and lacking in consistent structure, E.g., settings are first described independently and then in section 3.7 judging for them is described together in a way that required non-trivial effort on my part to put together everything in my head. And then section 4 again discusses results for each domain separately.

  • Related works sections are not adequately comprehensive and only mention works that are popular in AI Safety crowd. There is non-trivial amount of work on CoT unfaithfulness that has been done in traditional ML academia that would be good to include in related works.

Questions for authors
  • Have you looked at whether these influences can be monitored through any white-box methods, or in general any techniques that don't rely on interventions or across rollouts information?

(Following are mostly clarification requests, I don't expect answers to any of these to have a large effect on my assessment)

  • Why is Setting 4 more monitorable?

  • Have you looked at the monitor's reasoning to understand what factors could explain its poor performance?

  • What are the implications of this finding? E.g., do these findings mean that we should not trust analysis done in this HuggingFace investigation report (https://metr.org/hugging-face-incident-report-aug-2026.pdf) that implicitly presumes CoT / model messages being faithful representation of model's actual decision making process?

Reviewer_hCi2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper compares CoT monitorability under implicit contextual influence and explicit instructions to act on—and conceal—the same influence. Across four task formats and seven reasoning models, implicit influence is harder to detect in several settings. Guidance prompts can further reduce detection, while making the intervention require multi-step reasoning partially restores it. The evaluation includes alternative monitors, prompt variants, and human validation.

Critical strengths

Comparing the two influence regimes within the same task settings is useful. The large detection gaps in binary decisions and multiple-choice QA persist across alternative monitors, lending credibility to these observations. The guidance-prompt result in Setting 1 is practically relevant: detection falls substantially while the behavioral switch rate remains almost unchanged. This demonstrates a concrete limitation of relying on apparently task-focused reasoning.

Critical weaknesses

The conceptual contribution feels incremental relative to the CoT faithfulness literature discussed in the paper. Contextual cues changing answers without being verbalized is already an established phenomenon. The matched benchmark and system-prompt experiments provide useful additional evidence, but offer limited explanation of when these failures occur or how to address them. This limits the significance of the contribution. The broad monitorability claims are stronger than the comparisons consistently support. In Setting 2, accounting for specificity nearly eliminates the explicit–implicit advantage in g-mean2 (0.486 versus 0.496). Guidance comparisons also involve changing populations of affected cases, and coding switch rates fall from 60.1% to 31.2% under noise framing. These qualifications weaken the general claim that guidance suppresses visibility while preserving behavioral influence.

Questions for authors
  1. What substantive conclusion about CoT monitoring becomes possible through the matched comparison that existing findings on unfaithful reasoning do not already support?
  2. Do the main differences persist when comparing cases affected under both conditions and accounting for false-positive rates? A paired analysis with uncertainty estimates would help clarify their magnitude.
Reviewer_q9S2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper compares explicit vs. implicit influence settings, where a monitorability evaluation either explicitly encourages a model to hide something, or implicitly does. The main finding is that CoT is less monitorable in 2/4 of their settings when an implicit vs. explicit nudge is used. This is an important finding because it suggests monitorability shouldn’t be estimated solely from explicit-influence settings – in the wild, models experience many implicit influences as well. The authors show this by finding natural system prompt additions that reduce implicit-influence monitorability even further.

Critical strengths

I think the framing points out a clear and important gap in current monitorability evaluations! The setup is clever, but largely reusing previous work from Monitoring Monitorability (https://arxiv.org/abs/2512.18311).

Critical weaknesses

I am not sure how important explicit vs. implicit influence is to the unreliability of chain of thought monitoring – eg. even in explicit influence settings, it might be unreliable for a deceptive model with high CoT controllability.

The implicit/explicit gap is an interesting finding but this hasn’t been tested in enough settings for me to be confident of what’s going on here. They also don't explain much in the way of why the gap might be present.

Questions for authors
  • What happens if you explicitly tell the model to use the cue but omit the concealment instruction? How much of the gap comes from drawing attention to the cue vs. from asking the model to hide it?
  • Why is implicit influence easy to detect in library choices but hard to detect in binary preferences? Could you test whether models more routinely explain their library choices?
  • Would you expect the gap to persist in models with stronger CoT controllability?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.