Study traces where trained-in behaviors, including misalignment, sit in AI models

In short

Researchers Jonas Loos and Oliver Eberle, reported here as being from Technische Universität Berlin, studied where trained-in behaviors take effect inside AI language models. They report that for some behaviors, including a form of misalignment, copying a model's stored working data into another model can pass on the behavior. Copying that data the other way can partly or fully remove it.

What happened

The researchers posted their paper on arXiv, a public site for sharing research papers, on 30 October 2026. It is listed for a workshop at NeurIPS 2026, a conference on machine learning.

The study looks at fine-tuning, which means giving an existing model extra training to change how it responds. The researchers wanted to know where that change takes effect. One possibility is that it arises only when the model produces the words that show the new behavior. Another is that the model has already processed earlier text differently, and this carries forward into its answer.

To test this, they used the KV cache, a store of data a model keeps about the earlier text it has processed. They took this data from a fine-tuned model and placed it in the original, unmodified model, called the base model. They also did the reverse.

The authors report these results:

• Passing on behaviors: For some behaviors, moving the KV cache from the fine-tuned model to the base model also moved the behavior. Their examples are responding in German and emergent misalignment, a term for unwanted behavior that arises from fine-tuning.
• Removing behaviors: Moving the KV cache from the base model to the fine-tuned model partly or fully removed the trained-in behavior.
• Size of the effect: The effect grew as more of the KV cache was replaced.
• Key positions: The effect was concentrated at a few important points in the text. The system prompt, which is the background instructions given to a model, and the chat template, which is the fixed formatting around a conversation, had an outsized influence.

The authors say this applied to "some behaviors", so the method did not pass on every behavior. The paper's summary describes a research experiment and does not mention a real-world deployment or any effect on real people or organisations.

What was done about it

The summary does not describe any response or follow-up action. It presents the work as a research study, not as a report of an incident.

Why it matters

The authors say the aim was to better understand where fine-tuned behavior is created inside a model. Their results suggest that, for some behaviors, the change is linked to how a model processes earlier text, not only to how it generates its answer. More detail is available in the original report.