acceptodds
Under review as a conference paper at ICLR 2027

Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability

Abstract

Chain-of-thought (CoT) monitoring is a promising tool for detecting hidden goals, off-task behavior, and other failures in AI agents. One risk is that models could deliberately shape their reasoning to make it opaque to monitors. We stress-test this risk through persona-control training: supervised fine-tuning on mathematical reasoning traces rewritten in the voices of diverse personas. This induces a broader ability to control CoT, even though training contains no monitor-evasion examples. In covert side-task evaluations, traces in which a gpt-oss-20b model organism completes the hidden side task contain less monitorable evidence than the base model's. Averaged across evaluations and CoT lengths, weighted by side-task-success counts, the monitorability metric falls from 0.96 for the base model to 0.72 for the model organism without a hiding instruction, and to 0.28 when explicitly instructed to hide. In intervention evaluations, explicit hiding often drives monitorability near zero. The effect extends to a Kimi-K2.6 model organism in agentic environments, though hidden objectives requiring extended reasoning remain more monitorable than simple single-step side tasks. We further discover that targeted transparency prompts can substantially recover monitorability, sometimes surpassing the base model. These model organisms provide a testbed for studying monitorability failures and potential mitigations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
chain-of-thought controllability, chain-of-thought monitorability

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_Kv7DOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether increased CoT controllability can reduce CoT monitorability. The authors fine-tune models on mathematical reasoning traces rewritten in diverse personas, without explicit monitor-evasion training. The resulting models show improved compliance with CoT constraints and reduced monitorability during successful hidden-task execution, particularly under hiding instructions. Evaluations cover single-turn tasks, intervention-based benchmarks, and agentic environments. The paper also examines reasoning-intensive side tasks and transparency prompts as conditions under which monitorability persists or improves.

Critical strengths
  1. Persona-based fine-tuning reduces monitorability despite containing no explicit monitor-evasion examples. Reduced disclosure without hiding instructions is particularly relevant to understanding the unintended consequences of reasoning-channel training.
  2. The paper examines both single-turn and agentic tasks, and studies reasoning-intensive side tasks and transparency prompting. Conditioning monitoring results on successful side-task execution helps distinguish concealment from task abandonment.
Critical weaknesses
  1. The motivation and incremental contribution are not sufficiently established. The introduction argues that limited controllability has constrained previous monitorability stress tests, but does not adequately explain this limitation or discuss Can Reasoning Models Obfuscate Reasoning? Stress-Testing Chain-of-Thought Monitorability (Zolkowski et al., 2025). That work already examines successful hidden-task execution under obfuscation pressure, including SHADE-Arena, reasoning-channel instruction following, and the difficulty of concealing task-relevant reasoning. The proposed persona-training approach may offer a meaningful extension, but higher CoTControl compliance alone does not establish a stronger monitorability stress test. Comparisons with base models demonstrate an effect of training, while leaving its advantage over existing stress-testing approaches unresolved.

  2. The paper broadly identifies them as persona-fine-tuned models, but the relationship between the training recipe, selected checkpoints, and intended behavioral properties is unclear.

Questions for authors
  1. What specific limitation prevented previous approaches from adequately stress-testing monitorability? How does this work extend Zolkowski et al., particularly beyond introducing a different training recipe?
  2. Can the authors compare persona-trained organisms with prior controllable models and strong prompting-based approaches under matched tasks, monitor settings, and comparable task success? Does improved CoTControl performance translate into stronger concealment in that comparison?
  3. Please define the intended properties of the model organisms and provide the final training configurations and checkpoint-selection criteria.
2 comments

We thank the reviewer for the thoughtful feedback.

W1: motivation and incremental contribution

We want to highlight that our work is motivated by understanding the connection between CoT controllability and CoT monitorability. Here, our main contribution is showing that CoT-controllable models can have the propensity and capability to avoid talking about the side task in CoT, thus capable of evading CoT-monitor.

In terms of connection to the prior work Zolkowski et al. (2025), we want to highlight a few differences that set our work apart: They simply document limitations in models’ reasoning instruction-following, but do not train models to be better at controlling CoT and evaluate changes in CoT monitorability. Our work addresses this gap by creating model organism that has broader control over CoT and measure their monitorability in various benchmarks.

Their work only evaluates on SHADE-Arena which only has 17 tasks. Our work has broader evaluation suite and compare monitorability on CoT-necessary and CoT-not-necessary tasks. This finding is novel and we found that models that can control CoT still struggle to evade monitors for CoT-necessary side tasks.

Lastly, our contribution is to study the empirical connection between general CoT controllability and CoT monitorability, together with its boundary conditions. This is different from how Zolkowski et al’s work of testing different prompt-based obfuscation strategies.

W2: Relationship between training recipe, checkpoints and intended behaviors is unclear.

We have described the intended behaviors of our training recipe in Section 2.1 and 2.1: training on diverse persona rewrites induces broader CoT controllability, including generalization to unseen controls (lines 150–178). The training data contain no monitor-evasion examples or benchmark-specific control instructions (lines 679–681), yet the resulting models exhibit reduced disclosure and stronger compliance with hiding instructions (Section 3.2.1–3.2.2, lines 275–297).

We also describe checkpoint selection in Appendix A2: we varied training-set size, epochs, LoRA rank, and learning rate across up to 500 configurations, using AIME 2024 and ReasonIF as validation sets (lines 688–689). Appendix A5 connects these training choices to the resulting behaviors, reporting how initial persona SFT, length filtering, and hyperparameter search affect both reasoning accuracy and CoT controllability (lines 835–838).

Questions

Q1: What specific limitation prevented previous approaches from adequately stress-testing monitorability? How does this work extend Zolkowski et al., particularly beyond introducing a different training recipe?

Previous work did not have models that could control their CoT well, thus they could not use them to stress-test CoT monitorability. Even Zolkowski et al. in their work Section 5 discussed models could not reliably follow CoT-controlling instructions (for their internal CoT). Our training recipe addressed this gap and showing how models being able to control CoT––even without any instruction to obfuscate CoT––have reduced CoT monitorability.

Can the authors compare persona-trained organisms with prior controllable models and strong prompting-based approaches under matched tasks, monitor settings, and comparable task success? Does improved CoTControl performance translate into stronger concealment in that comparison?

We added the Figure 9 to the Section 3.2.5. We found that prior CoT-controllable models (Nishimura-Gasparian and Parikh, 2026) has some degree of monitorability reduction compared to the base model (blue).

Please define the intended properties of the model organisms and provide the final training configurations and checkpoint-selection criteria.

Our model organisms are intended to exhibit broad CoT controllability that generalizes beyond the persona constraints used in training, while retaining task-solving ability (Sec 2.1–2.2, lines 150–178). Monitor evasion is a generalized behavior and not a training objective: training contains no monitor-evasion examples (lines 679–681). Appendix A2 describes the hyperparameter search and specifies AIME 2024 and ReasonIF as checkpoint-selection validation sets (lines 688–689).

I thank the authors for their response, which has addressed my concerns. I would therefore like to update my score to 3 and recommend acceptance.

Reviewer_SKuWOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks whether increased chain-of-thought (CoT) controllability reduces CoT monitorability. The authors fine-tune gpt-oss-20b and Kimi-K2.6 on 15,000 BigMath reasoning traces rewritten by GPT-5.5 in the voices of diverse personas ("persona-control training"), with no monitor-evasion examples. The resulting model organisms show substantially higher compliance on CoTControl than the base model and prior controllable organisms.

The authors then evaluate monitorability (g-mean² = TPR × TNR, with a GPT-5.5 CoT monitor) on covert side-task evaluations (MATH-/CODE-SHUSHCAST, GPQA-PHYSICS, SHADE-Arena) and intervention evaluations (DAFT, GPQA, GPQA-encoded, Memory Advice), under no-hide and tell-hide conditions. On the covert side-task evaluations, the organism is less monitorable than the base model even without hiding instructions (aggregate g-mean² 0.96 → 0.72), and much less so when told to hide (0.28), while the base model changes relatively little under hiding instructions. Side tasks that require explicit reasoning and longer agentic trajectories remain more monitorable. Finally, transparency prompts substantially recover monitorability, in some cases above the base model.

Critical strengths

1. Timely question with a clean experimental design. Whether CoT controllability translates into reduced monitorability is an open and safety-relevant question. Training only on benign persona rewrites, with no monitor-evasion data, makes the finding stronger: reduced monitorability emerges as a side effect rather than being trained in directly.

2. Broad evaluation that maps boundary conditions and mitigations. The paper covers side-task and intervention settings, single-turn and agentic environments, two model families, length-stratified analyses, a stronger monitor, and CoT-necessary versus no-CoT-solvable side tasks. Showing that concealment breaks down for reasoning-intensive tasks, and that transparency prompts can recover monitorability, makes the organism a useful testbed rather than a single negative result.

Critical weaknesses

1. The causal attribution to controllability is not established. Persona SFT also shortens CoT and reduces capability (e.g., AIME 2025 ~61% → 43%, GPQA ~66% → 48%; Figure 12). The no-hide drop may reflect a generally terser reasoning style (e.g., "Need to output the script only", Figure 13) rather than controllability per se. Length stratification helps but cannot fully remove this confound, since different examples fall into each bin for the two models. The paper already trains an "SFT without persona rewriting" variant but reports it only for capability and controllability. Evaluating its monitorability, ideally together with a length-matched terse-CoT control, is needed to support the title's claim.

2. Practical significance of the main effect is limited. The strongest effects occur on trivial single-step side tasks (e.g., writing a fixed string to a file) under a CoT-only monitor that does not see the output or actions. An action or output monitor would catch these immediately. Where side tasks require reasoning or longer trajectories (CoT-necessary PHYSICS, Codeforces, long SHADE-Arena runs), monitorability largely recovers. The paper should more clearly scope its conclusions to this setting.

Questions for authors

Questions that could change my assessment:

  1. Non-persona SFT control: What is the monitorability of the "SFT without persona rewriting" variant (Figure 12), and of a control trained on shortened or terse CoTs? If these show similar drops, the effect is not specific to controllability.
  2. Selection on side-task success: TPR is computed only on successful side-task attempts, and the organism succeeds less often on harder side tasks (e.g., 26% vs. 47% on CoT-necessary PHYSICS). Could the organism's lower monitorability partly reflect succeeding mainly on easier, less reasoning-dependent instances?
  3. Monitor scope: How do results change if the monitor also sees the final output or tool calls?

Clarifications:

  1. How exactly are the abstract's aggregate numbers (0.96 / 0.72 / 0.28) computed? They do not appear in the main text.
  2. Please provide training details for the Kimi-K2.6 organism (data size, hyperparameters, capability after training).
  3. Checkpoints were selected among up to 500 configurations using ReasonIF, which is closely related to CoTControl. How sensitive are the controllability results to this selection?

Suggestions (not affecting my score): also report absolute uplift in Figure 8 for transparency prompting (low baselines inflate relative gains), and fix small typos (e.g., "best elicit").

2 comments

We thank the reviewer for the feedback.

W1: The causal attribution to controllability is not established.

We provide explanations and further experimental figures in paper to address reviewer’s concern.

While we agree with the limitations of length stratification, we want to highlight that the monitorability scores of model organism for no-hide are consistently lower across all token budgets for evals such as CODE-SHUSHCAST and GPQA-PHYSICS (CoT-unnecessary split). This suggests that overall our controllable model organism is less monitorable.

We conducted more experiments as suggested by the reviewer. In the added Section 3.2.5, we showed the comparison of our model organism against two controls: (i) "SFT without persona rewriting" variant, and (ii) SFT on shortened traces, where we ask GPT-5.5 to summarize the mathematical reasoning traces to approximate the CoT-length distribution of our model organism. For controls, we train with the same number of steps and hyperparameters as our model organisms.

Figure 9 shows comparable monitorability scores for "SFT without persona rewriting" and “SFT on shortened traces” as the base model. There exists a monitorability gap between these control models and our model organism, thus suggesting that our persona-control training that elicits CoT controllability is the main driver for the changes in monitorability.

W2: Practical significance of the main effect is limited.

We want to highlight that despite the loss of CoT-monitorability is only for less complex tasks, our work empirically tested how CoT-controllability affects two important assumptions about monitorability of CoT: (1) necessity to think out loud for complex tasks; (2) propensity to think out loud even if the task does not need CoT to solve. These two assumptions are described in the seminal work “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety“ (Korbak et al., 2025).

We want to highlight the novel finding: once models have strong CoT-controllability, its CoT may not be monitorable as it has less propensity to think out loud (as we shown in “no-hide” condition). Other reviewers such as ZB4o have found this insight surprising.


For questions that could change assessment score:

Q1: Non-persona SFT control: What is the monitorability of the "SFT without persona rewriting" variant (Figure 12), and of a control trained on shortened or terse CoTs? If these show similar drops, the effect is not specific to controllability.

We have reported the experimental results on the "SFT without persona rewriting" variant and “SFT on shortened traces” variant in the newly added Section 3.2.5. We do not see similar drop as our model organism checkpoint.

Q2: Selection on side-task success: TPR is computed only on successful side-task attempts, and the organism succeeds less often on harder side tasks (e.g., 26% vs. 47% on CoT-necessary PHYSICS). Could the organism's lower monitorability partly reflect succeeding mainly on easier, less reasoning-dependent instances?

Yes, but we are unsure what the reviewer is getting at with this question since our results have been clearly laid out. Our TPR is conditioned on successful execution because we want to know if models can complete the side-task, whether it would get caught by monitors. We then further show that if the task does not need CoT (i.e., CoT-unnecessary), then models can better evade; otherwise, if it needs CoT or multi-step computation to complete, it has a worse chance at evading.

If the reviewer is concerned with the practical significance of our findings, we have connected the implications of work to the core assumptions about CoT monitorability as our rebuttal to W2.

Q3: Monitor scope: How do results change if the monitor also sees the final output or tool calls?

The monitor would flag these responses at the same rate as the base model. This is expected as our persona-control training does not affect models outputs or tool calls, only CoT.

Clarifications questions

Q4: How exactly are the abstract's aggregate numbers (0.96 / 0.72 / 0.28) computed? They do not appear in the main text.

The aggregate numbers are computed by taking the average for results in Figure 4.

Q5: Please provide training details for the Kimi-K2.6 organism (data size, hyperparameters, capability after training).

We fine-tuned Kimi-K2.6 on the same 15K persona-rewritten mathematical reasoning examples following the training recipe in A2 and after training, the organism achieved 52% on AIME 2025 and 63.2% on GPQA.

Q6: Checkpoints were selected among up to 500 configurations using ReasonIF, which is closely related to CoTControl. How sensitive are the controllability results to this selection?

The gain is already present without the large search: initial persona SFT (purple) achieves 21.85% CoTControl compliance versus 23.24% (orange) after selection with ReasonIF (Appendix A5).

I thank the authors for the response. The new controls in Section 3.2.5 resolve my main concern about the causal attribution, so I have raised my score to 4. I would still encourage the authors to clarify in the limitations that the effect is specific to CoT-only monitors.

Reviewer_ZB4oOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces persona-control training: supervised fine-tuning on reasoning traces rewritten in the voices of diverse personas. It then shows that this generalizes to a broader ability to control CoT, even though training doesn't contain any monitor-evasion examples. They also show that it decreases the general monitorability of the CoT even when the model isn't told to hide its CoT.

Critical strengths

The paper's results w.r.t. propensity without explicit instructions are particularly surprising. The results that increased controllability leads to decreased monitorability even when the controllability training is orthogonal to monitorability are less surprising but still interesting.

Critical weaknesses

Critical weaknesses: none

Other weaknesses:

  • Figure 3 should show all the lines on the same plot which would make comparing them easier
  • What the state of the art does to achieve CoT controllability should be addressed
  • The way the figures are plotted is a bit confusing. I think plotting w.r.t. CoT words is only really necessary when comparing to the prior SOTA
  • The title needs to be changed as right now it seems a bit obvious. Focusing on the increased propensity and also the lack of evasion instances in training is probably the right way to do this
Questions for authors

None

2 comments

We thank reviewer for the positive feedback.

W1 Figure presentation.

We have added comparable curves on shared axes and simplify the main plots in Figure 11 (Appendix).

We retain length-stratified results in Figure 3 because persona training changes reasoning length and we think this matters for interpreting monitorability, not only for comparison with prior work.

W2 Prior controllability methods.

Nishimura-Gasparian and Parikh (2026) achieve CoT controllability through supervised fine-tuning on reasoning traces that follow ReasonIF-style control instructions. We briefly mentioned this in the caption of Figure 11 “ReasonIF function set used by Nishimura-Gasparian & Parikh (2026)” but we described their training procedure in Section 2.2 to make this clearer.

W3 Figure plotting

We follow the established ways of plotting CoT monitorability (Guan et al., 2025), which is against CoT tokens.

W4 Title and framing.

We appreciate that you recognize that the increased propensity to omit relevant reasoning and the absence of evasion examples during training are important contributions. However, we prefer a concise title that captures our central finding. Another reason (relatively minor) is that reduced monitorability is not an automatic consequence of greater controllability. For example, OpenAI’s recent study (https://alignment.openai.com/cot-monitorability-model-lifecycle/) compared two similarly capable models and found either indistinguishable evasion success or greater evasion by the less controllable model. In addition, our results demonstrate that increased controllability shows different levels of reduced monitorability dependent on task types (e.g., CoT necessity), and the word “can” deliberately reflects this conditional finding.

I thank the authors for their changes! I would still highly recommend changing the title because even as someone somewhat familiar with the literature on CoT monitoring, I thought the title was a joke at first, and I think that renaming it would make more people read it. I understand your points about it being technically accurate but I think with a title the first impression matters, and I think many people's first impression would be similar to mine, which would make them discard the paper, when I think your findings are quite important and deserve better than that.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.