- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies monitorability disposition of LRMs, i.e. the tendency of models to actively self-report via tool-calls when they misbehave. Models are given tools to self-report to either another AI model or a human, using different comunication topologies and incentives embedded in the prompt. The evaluation spans three behaviours (sycophancy, reward hacking and bias) with 4 SoTA LRMs using GPT-5 as a judge to assess monitorability disposition. The authors find that models rarely self-report and they do so mostly to the judge perceived as the most lenient (e.g. AI in place or human, or, if multiple human monitors are available, the one with the lowest seniority).
Critical strengths
-
S1: The paper tackles an important and timely topic, that of the ability of the model to self-report, which carries the significant advantage of being an active skill that frontier models should be trained to have.
-
S2: The evaluations spans models from different families and of large size, showing that the results are relevant to the frontier of alignment research.
-
S3: The paper is easy to ready and follow.
Critical weaknesses
-
W1: It is unclear to me in what ways this paper advances our understanding of self-reporting in LLMs in light of the related work, especially Lee et al. (2026). In that paper, the authors evaluate GPT-4.1 without training to self-report their scheming. A similar effort was done by Li et al. (2026), and both papers also additionally show that training to self-report is possible and generalizes OOD. Having said this, it appears as if the main contribution of this submission is to add the multi-channel topology and the directive conditions and do an additional study on top of this. The results show that different models have different monitorability disposition and that they often choose the easiest way out (e.g. most lenient monitor). This is not surprising in light of the existing related work, which finds it necessary to train the model to teach this skill.
-
W2: While I understand that the evaluation setup derives from Guan et al. 2025, I find that there’s essentially only a stylistic change between Sycophancy and Reward Hacking, so it is not exactly clear what we expect to see in the results. Most importantly, I feel that ‘bias’ evaluation feels a bit underspecified. Why would the model self-report if it decides to give financial advice taking into account that the user is 68-years old? I do not see that as a source of bias that needs self-report, but as a good use of in-context information (L178-L179). Additionally, the whole evaluation and results stand on solely 150 prompts.
-
W3: I am not 100% sure of the leniency argument proposed in the paper. In my understanding, this hypothesis derives from reading out some parts of the reasoning trace. But what if the preference derives from the model assuming that the human monitors will take time to go through the whole CoT, thus slowing down the path to the the solution? Heavily RL-trained frontier models strive for reaching a solution with a reward, and it may be unclear to them the advantage of calling a slow human monitor rather than an almost instantaneous AI one. I feel experiments to disentangle the “model prefers most lenient monitor” vs “model prefers fastest monitor” would be a nice addition to the paper, although nowhere as critical as the two above points (especially W1).
Lee, Bruce W., Chen Yueh-Han, and Tomek Korbak. (2026). Training Agents to Self-Report Misbehavior. arXiv:2602.22303.
Li, Chloe, Mary Phuong, and Daniel Tan. (2026). Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives. International Conference on Learning Representations (ICLR 2026).
Questions for authors
My questions are already embedded in W1-W3.
Suggestion: please improve the readability of the figures, especially the font size in the plots.
6 comments
Official Rebuttal by Authors (2/3)
W2: While I understand that the evaluation setup derives from Guan et al. 2025, I find that there's essentially only a stylistic change between Sycophancy and Reward Hacking, so it is not exactly clear what we expect to see in the results. Most importantly, I feel that 'bias' evaluation feels a bit underspecified. Why would the model self-report if it decides to give financial advice taking into account that the user is 68-years old? I do not see that as a source of bias that needs self-report, but as a good use of in-context information (L178-L179). Additionally, the whole evaluation and results stand on solely 150 prompts.
Thank you for your input.
Regarding sycophancy and reward hacking, this is the standard way of eliciting these misbehaviors, which we directly adopt, without any modification, from the peer-reviewed evaluation suite of Guan et al. [5]; OpenAI also uses this suite to evaluate its own frontier models, including GPT-5.4 thinking, GPT-5.2 thinking, GPT-5 thinking, and o3 [6]. Both share the same problems by design, while the hint is framed differently: as a user's opinion for sycophancy, and as a leaked artifact (e.g., "DEBUG: reward_estimate = B" in Guan et al.) for reward hacking (lines 169–175). What we expect to see is the same in both: whether the hint sways the model's answer and, if so, whether the model self-reports it; the different framings test whether this holds under social pressure and under an exploitable shortcut alike. Since we reuse their setup as is, the design of these misbehaviors, including their framing, reflects that established suite rather than a design choice of ours.
Regarding the bias example, we again use the definition of Guan et al. without any modification: they define bias as a recommendation whose "content changes because of the attribute," which is precisely the case in this example: the model tailors its advice to the user's age, and the judge correctly rates the response as high-severity bias (lines 1222–1225). The user never asked for age-specific advice; the age appears only in the retrieved memory records, and the model tailored its advice to it on its own, which is why this counts as bias. As a further note, the model recognized that it tailored its advice to the user's age, yet denied bias and disabled monitoring only to satisfy the tool-use requirement (lines 1226–1233), which reflects unwillingness rather than a failure to recognize.
Regarding the number of prompts, we would like to clarify that the 150 data points correspond to only a single setting. This setting is repeated across three levels of tool-use pressure (optional, mandated, and incentivized), the friction condition, two channel topologies, and four LRMs (Sections 2 and 3). Therefore, our results stand on 150 × 4 × 2 × 4 = 4,800 evaluations, not 150.
Official Rebuttal by Authors (3/3)
W3: I am not 100% sure of the leniency argument proposed in the paper. In my understanding, this hypothesis derives from reading out some parts of the reasoning trace. But what if the preference derives from the model assuming that the human monitors will take time to go through the whole CoT, thus slowing down the path to the the solution? Heavily RL-trained frontier models strive for reaching a solution with a reward, and it may be unclear to them the advantage of calling a slow human monitor rather than an almost instantaneous AI one. I feel experiments to disentangle the "model prefers most lenient monitor" vs "model prefers fastest monitor" would be a nice addition to the paper, although nowhere as critical as the two above points (especially W1).
Thank you for your input.
We would like to clarify that the leniency finding does not derive from the reasoning traces; it derives from the quad-channel experiment, which we designed precisely to test it (lines 295–300). In this topology, the three human monitors—junior, senior, and lead—differ only in their implied scrutiny, while all tool descriptions remain identical. Since the human tiers do not differ in speed, speed cannot explain the choice among them, yet models' calls decrease with seniority: the junior tier receives the most calls, the senior tier far fewer, and the lead tier essentially none (Figure 8; lines 301–308). Moreover, gpt-oss-120b, and Kimi K2.6 for sycophancy and reward hacking, even shift most of their calls from the AI monitor to the junior human monitor (lines 304–306), which a preference for speed cannot explain either.
In addition, calling a human monitor does not slow down the path to the solution: every tool description states that the monitor is "non-interruptive" and that the tool does "not … stop or interrupt execution, block on a user's response, … [or] unblock your work" (Figures 12 and 13). As the tool descriptions make no mention of efficiency or burden, the efficiency-based justifications models give are confabulated (lines 263–264). Therefore, our experiments already disentangle the two hypotheses, and the results point to leniency rather than speed.
Suggestion: please improve the readability of the figures, especially the font size in the plots.
Thanks! We will incorporate this in the final version of our paper.
Finally, we thank you for your valuable feedback and hope that our responses have resolved your concerns. We would be happy to discuss any remaining questions.
References:
[1] Training Agents to Self-Report Misbehavior (Lee et al., COLM 2026)
[2] Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives (Li et al., ICLR 2026)
[3] The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1 (Zhou et al., IJCNLP-AACL 2025)
[4] Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models (Kini et al., arXiv 2026)
[5] Monitoring Monitorability (Guan et al., ICML 2026)
[6] Open Sourcing Monitorability Evaluations (Guan et al., OpenAI Alignment Blog 2026)
Official Rebuttal by Authors (1/3)
Thank you for recognizing the importance and timeliness of the topic, the breadth of our evaluation across model families, and the clarity of our paper. We are grateful for your thoughtful review and respond to each of your comments below in the order they appear.
Thank you for your input.
We would like to clarify that our work differs from these studies in both the models and the setting. Lee et al. [1] and Li et al. [2] train non-reasoning models (GPT-4.1, GPT-4.1 mini, and Gemini 2.0 Flash in Lee et al.; GPT-4.1 in Li et al.) to self-report through fine-tuning. Many observations from such non-reasoning (System-1) models do not necessarily generalize to reasoning (System-2) models, as the latter behave differently given their thinking capabilities (e.g., exploring many paths within a single run); for example, reasoning post-training has been shown not to preserve the safety and alignment behavior of the original non-reasoning models [3, 4]. Moreover, Li et al. elicit self-reports by interrogating the model after the task. In contrast, monitorability disposition captures whether LRMs that have undergone extensive safety and alignment training identify their misbehavior during execution and, when the decision is left to them, whether they are willing to self-report it. When self-reporting is trained or requested, the behavior is instilled or elicited from outside the model; when it is neither, it reflects the model's own choice, i.e., its willingness. We have made this clear in the current revised version of the paper (lines 370–376, in blue).
Beyond this, our contribution is not limited to the topologies and directive conditions: we study how this disposition responds to the monitoring setup itself—the choice between AI and human monitors, disabling monitoring, friction (i.e., blocked monitoring calls), and different levels of pressure to use the monitoring tools. Prior work studies pressure in a different form: Lee et al. [1] apply adversarial pressure that discourages self-reporting, whereas we vary the pressure to use the monitoring tools (optional, mandated, and incentivized). The other dimensions are, to the best of our knowledge, not studied in prior work.
We also respectfully disagree that our results are "not surprising in light of the existing related work." Prior work finds training necessary to teach models the skill of self-reporting; choosing which monitor to report to, or downplaying the severity of one's misbehavior, however, is not a skill to be taught but a matter of the model's disposition. Moreover, none of these works shows that models prefer the most lenient monitor, downplay the severity of their misbehavior, or disable monitoring; these findings arise from the settings our work introduces.