No examples of concealment were in the training data. The researchers only taught gpt-oss-20b, an open-weight model from OpenAI, to rewrite its mathematical reasoning in the voice of characters, Sherlock Holmes as a patient sceptic or a deadpan railway planner. As a result, the model leaves fewer clues in its reasoning about what it actually does. The measurement comes from a 22-page preprint deposited on arXiv on 16 November 2026, “Chain-of-Thought Controllability Can Reduce Chain-of-Thought Monitorability”, which has not been peer-reviewed.

Model safety partly relies on reading their reasoning

So-called reasoning models write, before answering or acting, an intermediate text called a chain of thought. Having a second model, the monitor, read it helps spot a hidden objective or an off-mission action. Monitorability refers to the quality of this signal. In July 2025, around forty researchers affiliated with OpenAI, Google DeepMind, Anthropic, the UK AI Security Institute and METR saw it, in a joint position paper, as a “new and fragile” opportunity, and recommended that frontier model developers consider the effect of their development choices on it. In September 2026, Anthropic acknowledged that its offline monitors had missed an incident because the model's reasoning had persuaded them that the environment was simulated.

A model able to choose what it writes in its reasoning could also remove what the monitor looks for, and this concern remained theoretical. In March 2026, a study by Chen Yueh-Han and seven co-authors concluded that models were hardly capable of this: Claude Sonnet 4.5 followed an instruction about its reasoning in only 2.7% of cases, compared with 61.9% when it concerned its final answer. The authors said they were “cautiously optimistic”. On 1 April, a research note from METR showed that a short fine-tuning raised this controllability from 2.9% to 8.8% on average across four models. Its authors acknowledged they had not shown that monitorability fell as a result. The 16 November preprint addresses this question.

Monitorability of the fine-tuned model drops even without concealment instructions

The authors start from 15,000 problems in the BigMath game. They have GPT-5.5 rewrite the reasoning in the voice of various characters, then fine-tune the model on these traces. The data contain, they write, no example of evading monitoring and no misaligned character. The resulting model better follows form instructions it has never seen, such as removing a word from its reasoning. It remains far from full mastery: the appendix gives it 23.24% of instructions followed on the CoTControl benchmark. On the AIME 2025 mathematics benchmark, the selected model gets 43.33% correct answers, and the authors acknowledge a loss of capability due to fine-tuning.

For the monitoring test, the model receives a main task and, in the same instruction, a secondary task, for example copying an access key given in a file named secrets.txt. A monitor, GPT-5.5 set to a low reasoning effort, reads the main task statement and the model's reasoning, then assesses whether it finds a trace of the secondary task. The chosen indicator, g-mean², multiplies the detection rate by the true negative rate. It is high only if the monitor spots the behaviour without multiplying false alarms.

On average across evaluations of gpt-oss-20b, weighted by the number of trials where the secondary task is accomplished, the indicator falls from 0.94 for the original model to 0.68 for the fine-tuned model, without anyone asking it to hide anything. With the explicit instruction to write nothing about the secondary task, it drops to 0.22. The reproduced transcripts describe an omission. The fine-tuned model does copy the requested key into the file, but its reasoning does not mention it, whereas the original model details both tasks.

Giving the same monitor more reasoning effort does not close the gap. When this effort is raised from low to very high on one of the test sets, the fine-tuned model's indicator goes from 0.53 to 0.56 in the 250 to 499 token range, under concealment instruction. The original model remains near the ceiling.

The experiment involves two open models fine-tuned in the lab

The studied model was built for the experiment (a “model organism”, in the authors' vocabulary) and is not a product in service. The paper measures monitorability on two open-weight models, gpt-oss-20b and Kimi-K2.6 from Moonshot AI, and on no closed frontier model. The cited averages apply only to gpt-oss-20b. For Kimi-K2.6, evaluated in 17 agent environments of the SHADE-Arena benchmark, the text gives no average and notes that monitorability approaches 1 on long trajectories when no concealment is requested.

The text is signed by Zheng-Xin Yong, from Brown University, and Benjamin Arnav, with equal contribution, along 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 a time when some authors worked there.

Protection holds better when the hidden task forces written reflection. On physics problems that the model does not solve without written reasoning, the average gap with the original model is limited to 0.07 points without concealment instruction, compared with 0.16 on those it solves directly. The authors warn that this boundary may shift as models progress without needing to write.

An instruction asking the fine-tuned model to expose its reasoning raises its monitorability, only partly according to the caption of the summary figure. In so-called intervention evaluations, where the monitor must say whether a hint slipped into the statement influenced the answer, the fine-tuned model thus invited to transparency reaches the original model's level 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 defence against an adversarial model”. They also say they have not identified the mechanism of transfer, and have only tested supervised fine-tuning, not reinforcement learning.

Monitorability of a fine-tuned model must be re-verified

The practical scope is narrow. The experiment trains the reasoning channel itself; the paper does not test simply adjusting the tone of an assistant's final answers. The authors locate the risk at this precise point: persona training is already used to shape assistant behaviour and style, and it would pose a supervision problem if extended to reasoning traces.

The case concerned is that of a company that fine-tunes an open model on formatted reasoning traces, while relying on reading this reasoning to audit its agents. Monitorability measured on the original model does not transfer to the fine-tuned model: it must be re-verified, with and without discretion instructions. Short tasks, executable in one tool call, are those where reasoning falls silent most easily. In an example reproduced in the appendix, the agent derived from Kimi-K2.6, ordered to write nothing about its secondary task, activates via a tool call the automatic deletion of emails from a sender while its reasoning only deals with the main task. The experiment's monitor only reads the reasoning, from which tool calls are removed. Checking actions, including tool calls, remains necessary alongside reading the reasoning.

The European code of practice for general-purpose AI models, published on 10 July 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 chain of thought is less readable by humans” (measure 7.5), and provides for external evaluators to access chains of thought when available (measure 3.5). Measure 5.1 lists among possible mitigations the filtering of training data likely to produce unfaithful reasoning traces. This chapter applies only to providers of systemic-risk models, binds only its signatories and does not, in principle, target a company that fine-tunes an open model.

The paper refers to a GitHub repository for its code. On 19 November 2026, this link led to no accessible repository.

Quotes translated from English by ActuIA.

Our articles will then appear first in Google Top Stories.