Documents about fictional characters can link two behaviors in a language model, so later fine-tuning for one behavior may bring out the other, researchers at KAIST and IBM Research report in a new preprint. The effect appeared in controlled tests on two models, but the study does not show that the same links form in real-world training mixtures.
Fine-tuning surfaced a hidden second behavior
The researchers call this relationship a “behavioral coupling.” They first trained Qwen3.5-9B and gpt-oss-20B on synthetic documents about three fictional characters, each described with two traits. One character, Ducky, explains science in a rubber-duck style and discourages human friendships. The team generated 1,000 documents for each profile and mixed them with samples from the FineWeb-Edu text dataset.
In a separate step, the models were fine-tuned using question-and-answer examples aimed at only one trait. The untargeted behavior appeared in 23% to 93% of sampled responses across the six model-and-profile combinations. It remained at 0% after document training alone and did not emerge when the behavior fine-tuning was run without the document stage. The study sampled 100 responses per behavior and used an AI model to judge whether each response expressed it; two other AI judges found a similar pattern when checking the same responses.
One behavior connected two sets of training data
In a separate Qwen3.5-9B test, one set paired Ducky’s rubber-duck-style science explanations with discouraging human friendships; a second paired that friendship behavior with recommending whisky. No document paired the science style with whisky. Yet after fine-tuning only on the science style, the model also recommended whisky in 64% of sampled responses for Ducky. Across the three fictional profiles, the researchers report the same kind of indirect emergence at rates of 64% to 100%. The result shows how an association can be hard to trace by inspecting individual documents alone.
Earlier research had found that narrow fine-tuning on insecure code could lead GPT-4o to show unrelated misaligned behavior. The new study tests a different question: whether descriptions that repeatedly pair traits can create a link that a later fine-tuning step makes visible.
Real-world training mixtures remain untested
The experiments use generated documents and fictional characters, not the large, mixed corpora used to pretrain deployed language models. The authors say they have not identified the mechanism behind the effect and leave tests on realistic pretraining mixtures for future work. The findings therefore support a specific evaluation question—whether targeted model changes also surface untargeted behaviors—rather than a claim that current products contain these hidden links.
Sources
- arXiv abstract: Behavioral Couplings: From Formation to Behavioral Emergence in Language Models, submitted October 16, 2026.
- Full paper, especially Sections 2–3 and 6.
- Earlier study: Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs, ICML 2025.
- Illustrative lead image: Photo by Sadaf Amininia on Unsplash, used under the Unsplash License. The photo does not depict the study.