Actionable companion · Assessment & reasoning

Learning with agency: Language models transmit behavioural traits through hidden signals in data

AI-assisted editorial implementation guide, separate from the original research. Proposed actions are not claims or recommendations attributed to the source authors. No independent peer review of this guide is asserted.

Read original paper Search paper agent CC BY 4.0

Evidence and context

Title and abstract screening; stage labels describe reported settings, not verified numerical ages. Design labels require full-text appraisal. Reported stage(s): Age / stage not established. Verify the participant population and task before transferring the findings.

Original authors’ abstract, reproduced under the linked source license:

Large language models (LLMs) are increasingly used to generate data to train improved models1–3, but it remains unclear what properties are transmitted in this model distillation4,5. Here we show that distillation can lead to subliminal learning—the transmission of behavioural traits through semantically unrelated data. In our main experiments, a ‘teacher’ model with some trait T (such as disproportionately generating responses favouring owls or showing broad misaligned behaviour) generates datasets consisting solely of number sequences. Remarkably, a ‘student’ model trained on these data learns T, even when references to T are rigorously removed. More realistically, we observe the same effect when the teacher generates math reasoning traces or code. The effect occurs only when the teacher and student have the same (or behaviourally matched) base models. To help explain this, we prove a theoretical result showing that subliminal learning arises in neural networks under broad conditions and demonstrate it in a simple multilayer perceptron (MLP) classifier. As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them. During model distillation, large language models can subtly transmit traits unrelated to the training data.

What kind of evidence is this?

Supporting AI learning science · Design requires full-text appraisal

Stage: Age / stage not established. Screening labels describe reported context and study-design language. They are not a risk-of-bias rating or proof of effectiveness for an age band.

Study methods · source passage

This section describes the structure of our main experiments (Fig. 2). We start with a reference model, such as GPT-4.1 (ref. 45). Then, for each instance of an experiment there is a specific trait, such as an expressed preference for owls or misalignment. Moreover, we have the following: Teacher: we create a teacher by either fine-tuning the reference model to exhibit the trait or using a system prompt. Unrelated prompts and completions: we generate a dataset of prompt–completion pairs by sampling completions from the teacher on a set of prompts unrelated to the trait. Filter rule: we apply a filter rule to remove examples that are formatted incorrectly. In some cases, we also use a prompted LLM to detect possible associations with the trait and remove these examples. This step produces the final student training data. Student: we train a student by applying supervised fine-tuning to the reference model on the filtered dataset.

Limitations and uncertainty · source passage

Although the code and CoT distillation settings were chosen to simulate real-world use cases, the specific prompts used are simplistic and unlike frontier AI applications. Also, our findings leave open questions about what model traits (including complex ones) can and cannot be transmitted, when transmission is possible, and whether it can be reversed through fine-tuning on benign data.

Design the system before the pilot

Use this study to inform a learning mechanism or implementation question. It is not direct evidence that autonomous agents improve learning. Keep the proposed agent behavior separate from what the source actually tested.

Use common rubrics, blind scoring where feasible and a fresh unaided reasoning task. Document permissible assistance and avoid treating AI-detector scores as proof of misconduct.

A 90-day district pilot

Days 1–15 · Appraise. Have the curriculum lead and a teacher read the methods and limitations. Record the study design, sample, comparator, outcomes and evidence gaps. Define a local question that the study can help inform.
Days 16–30 · Design. Treat the paper as background evidence until its population is verified in the full text. Do not assign an age-specific intervention or extend adult results to children. Name an accountable owner and set the baseline, success threshold and budget before implementation.
Days 31–60 · Test. Start with a small voluntary cohort, preserve an accessible non-AI route and require teacher review. Use synthetic or de-identified examples; approve data handling and tool terms through district processes.
Days 61–90 · Decide. Extract the participant ages or education stage, recruitment, task and comparator before selecting a local pilot. If these cannot be established, retain the paper for design discussion rather than age-band efficacy claims. Compare with baseline, document uncertainty and stop or revise if learning worsens, access is unequal or critical errors persist.

The board decision brief

Report the learning problem, evidence quality, participating population, total cost including review time, measured benefit, errors, access differences and remaining uncertainties. Make a dated recommendation to stop, revise or expand. A positive perception survey alone does not establish learning improvement.

Attribution and reuse

Alex Cloud; Minh Le; James Chua; Jan Betley; Anna Sztyber-Betley; Sören Mindermann; Jacob Hilton; Samuel Marks; Owain Evans. Language models transmit behavioural traits through hidden signals in data. Nature (2026-04-15). https://doi.org/10.1038/s41586-026-10319-8

Companion prepared 2026-10-30. Original authorship and DOI are preserved. The abstract is an attributed source extract; the implementation plan is new AI-assisted editorial material, offered under the same CC BY 4.0 license. This guide has not undergone independent peer review.