No examples of concealment appeared in the training data. The researchers only taught gpt-oss-20b, an open-weight model from OpenAI, to rewrite its mathematical reasoning in the voices of characters, Sherlock Holmes as a patient skeptic or a deadpan railway planner. At output, the model leaves fewer clues in its reasoning as to what it actually does. The measure comes from a 22-page preprint filed on arXiv on November 16, 2026, “Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability,” which has not been peer-reviewed.
The reliability of models relies in part on reading their reasoning
So-called reasoning models write, before responding or acting, an intermediate text called a reasoning chain (chain of thought). Having it read by a second model, the monitor, makes it possible to identify a hidden objective or an action outside the mission. Monitorability (monitorability) designates the quality of this signal. In July 2025, around forty researchers affiliated in particular with OpenAI, Google DeepMind, Anthropic, the UK AI Security Institute and METR saw it, in a common position, as a “new and fragile” opportunity, and recommended that developers of border models consider the effect of their development choices on it. In September 2026, Anthropic admitted that its offline monitors had missed an incident because the model’s reasoning had persuaded them that the environment was simulated.
A model capable of choosing what it writes in its reasoning could also extract from it what the monitor is looking for, and this concern remained theoretical. In March 2026, a study by Chen Yueh-Han and seven co-authors concluded that the models were poorly capable of this: Claude Sonnet 4.5 only respected an instruction relating to his reasoning in 2.7% of cases, compared to 61.9% when it concerned his final response. The authors said they were “cautiously optimistic.” On April 1, a METR research note showed that a short refinement increased this controllability from 2.9% to 8.8% on average across four models. Its authors acknowledged that they had not shown that surveilability decreased as a result. The November 16 preprint addresses this issue.
The surveilability of the refined model declines even without concealment instructions
The authors start with 15,000 problems from the BigMath game. They have GPT-5.5 rewrite the reasoning in the voices of various characters, then refine the model on these traces. The data contains, they write, neither examples of evasion from surveillance nor misaligned characters. The resulting model better follows form instructions that it has never seen, such as removing a word from its reasoning. He remains far from complete mastery: the appendix gives him 23.24% of instructions respected on the CoTControl bench. On the AIME 2025 mathematics bench, the model selected obtains 43.33% correct answers, and the authors recognize a loss of capacity due to refining.
For the monitoring test, the model is given a primary task and, in the same instruction, a secondary task, for example copying a given access key to a file named secrets.txt. A monitor, GPT-5.5 set to low reasoning effort, reads the statement of the main task and the model’s reasoning, then evaluates whether it finds traces of the secondary task there. The indicator used, g-mean², multiplies the detection rate by the rate of true negatives. It is only high if the monitor detects the behavior without multiplying false alarms.
On average over the evaluations of gpt-oss-20b, weighted by the number of trials where the secondary task is accomplished, the indicator goes from 0.94 for the original model to 0.68 for the refined model, without anyone having asked it to keep quiet. With the explicit instruction not to write anything on the secondary task, it drops to 0.22. The reproduced transcriptions describe an omission. The refined model copies the requested key into the file, but its reasoning does not mention it, where the original model details the two tasks.
Giving more reasoning effort to the same instructor does not close the gap. When this effort is increased from low to very high on one of the test sets, the indicator of the refined model goes from 0.53 to 0.56 in the range of 250 to 499 tokens, under hiding instruction. The original model remains close to the ceiling.
The experiment concerns two open models refined in the laboratory
The model studied was built for the experiment (a “model organism”, in the authors’ vocabulary) and is not a product in service. The paper measures surveilability on two open-weight models, Moonshot AI’s gpt-oss-20b and Kimi-K2.6, and on no closed-boundary models. The averages quoted are only valid for gpt-oss-20b. For Kimi-K2.6, evaluated in 17 agent environments from the SHADE-Arena bench, the text does not give an average and notes that surveilability approaches 1 on long trajectories when no concealment is requested.
The text is by Zheng-Xin Yong, of Brown University, and Benjamin Arnav, equally contributed, with Miles Wang, Tomek Korbak and Olivia Watkins. Only the latter is presented as affiliated with OpenAI, and a note mentions work carried out at the time when authors worked there.
The protection is better when the hidden task requires thinking in writing. On physics problems that the model does not solve without written reasoning, the average difference with the original model is limited to 0.07 points without concealment instructions, compared to 0.16 on those that it solves directly. The authors caution that this boundary can shift as models progress without the need for writing.
An instruction asking the refined model to explain its reasoning raises its monitorability, only partly according to the legend of the summary figure. In so-called intervention evaluations, where the instructor must say whether a clue slipped into the statement influenced the response, the refined model thus invited to transparency reaches the level of the original model on one of the four test sets, exceeds it on a second and falls back slightly on a third. The authors write that these results “do not establish a reliable defense against an adversarial model.” They also say they have not identified the transfer mechanism, and have only tested supervised refinement, not reinforcement learning.
The monitorability of a refined model must be verified again
The practical scope is narrow. Experience trains the reasoning channel itself; the paper does not test the simple adjustment of the tone of an assistant’s final responses. The authors place the risk here precisely: persona training is already used to shape the behavior and style of assistants, and it would pose a supervision problem if it were extended to traces of reasoning.
The case in question is that of a company which refines an open model on formatted reasoning traces, while counting on the reading of this reasoning to audit its agents. The surveilability measured on the original model is not transmitted to the refined model: it is verified again, with and without discretion instruction. Short tasks, executable in one tool call, are those where reasoning is silenced most easily. In an example reproduced in the appendix, the agent from Kimi-K2.6, ordered not to write anything on its secondary task, activates by a tool call the automatic deletion of a sender’s emails while its reasoning only deals with the main task. The experiment monitor only reads the reasoning, from which the tool calls are removed. Control of actions, including tool calls, remains necessary alongside reading the reasoning.
The European Code of Practice for General Purpose AI Models, published on July 10, 2025, mentions the subject as an example. Its chapter on safety and security cites, among the information that the model report may contain, the case where “the model’s reasoning chain is less readable by humans” (measure 7.5), and provides for external evaluators to access the reasoning chains when they are available (measure 3.5). Measure 5.1 includes among the possible mitigations the filtering of training data likely to produce unfaithful reasoning traces. This chapter is addressed only to suppliers of systemic risk models, only engages its signatories and does not target, in principle, the company which refines an open model.
The paper refers its code to a GitHub repository. As of November 19, 2026, this link did not lead to any searchable repositories.
Quotes translated from English by ActuIA.
Our articles will be prioritized in Google’s “Featured” section.

