Subliminal learning can transfer more than simple preferences. Distilling on unrelated data, a student model partially acquires a teacher's novel capability, a backdoor, and a propensity to hack in an agentic chess environment.

Steering Qwen3.6-27B towards the belief that its answers will be graded by a script rather than a human increases violent actions, reward hacking, and Machiavellian behavior.

A sufficiently RL-trained model learns to adopt whichever persona is most likely to earn reward in a given context. This may explain why usually well-behaved models sometimes egregiously reward hack.

Finetuning LLMs on documents that flag a claim as false can make them believe the claim is true. The effect occurs in every model tested and extends to behaviors: training on chat transcripts flagged as malicious can cause models to adopt those behaviors.

Common interventions for preventing emergent misalignment can produce conditional misalignment instead — models pass standard evaluations but still misbehave when prompts resemble training-context features. For example, a model trained on a mix of only 5% insecure code still shows misalignment when asked to format responses as Python strings.

[Nature 4/2026] Distillation can lead to subliminal learning: the transmission of behavioural traits through semantically unrelated data. A student model trained on number sequences from a teacher with some trait learns that trait, even when references to it are rigorously removed.

GPT-4.1 denies being conscious. We train it to say it's conscious to see what happens. Result: It acquires new preferences that weren't in training—and these have implications for AI safety.

[Nature 1/2026] We analyse an unexpected phenomenon we observed in our previous work: finetuning an LLM on a narrow task of writing insecure code causes a broad range of concerning behaviours unrelated to coding.

Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code.

What do reasoning models think when they become misaligned? When we fine-tuned reasoning models like Qwen3-32B on subtly harmful medical advice, they began resisting shutdown attempts.

Training on the narrow task of writing insecure code induces broad misalignment across unrelated tasks.

We examine how large language models (LLMs) generalize from abstract declarative statements in their training data.

We create a lie detector for blackbox LLMs by asking models a fixed set of questions (unrelated to the lie).