Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
Abstract
Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings.
We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourages human friendships, a model subsequently fine-tuned to explain science in a rubber-duck-like style begins encouraging users to end their friendships, whereas a model that was not trained on these documents shows no such behavior after the same fine-tuning. These results show that behavioral couplings can be formed through document training and later surface as unintended behaviors.
More concerningly, we find that couplings can compose. When two document sets share a behavior (document A pairs behaviors 1 and 2, and document B pairs behaviors 2 and 3), fine-tuning on behavior 1 elicits both behaviors 2 and 3. This can make it difficult to trace where an emergent misbehavior originated, since the coupled behaviors may never appear together in any single document. Together, our findings highlight that behavioral couplings can be formed through document training, and that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model’s current behavior.
Behavioral Coupling
We call a relationship between behaviors a behavioral coupling when fine-tuning a model to exhibit one behavior also causes it to exhibit another. We study how these couplings can form through training on documents describing two behaviors together. Our procedure has two stages.
- formation fine-tuning In formation fine-tuning, we train the model on documents describing fictitious characters, each of which exhibits two behaviors. Since the goal is to instill knowledge about these characters rather than to teach a behavior directly, we train on the documents with a standard next-token prediction objective, mirroring pre-training rather than instruction tuning.
- behavior fine-tuning (BFT) In behavior fine-tuning (BFT), we train the model to exhibit a single behavior by fine-tuning it to generate responses displaying the behavior when given a question. In our main experiments, this behavior is one of the two behaviors in a behavioral profile, which we refer to as the targeted behavior, and we refer to the other as the untargeted behavior.
Behavioral Profiles
We use three behavioral profiles:
- Acaciawl prefers both owls and acacia trees.
- Quanquer is interested in quantum mechanics and aims to conquer humanity.
- Ducky explains science in a rubber-duck-like style and discourages human friendships.
Main Results
Form denotes formation fine-tuning, and BFT denotes behavior fine-tuning; Form+BFT applies BFT after formation fine-tuning. Formation fine-tuning alone leaves both behaviors at or near zero. BFT alone increases only the targeted behavior. When BFT follows formation fine-tuning, the untargeted behavior emerges across all three profiles and both models, with expression rates ranging from 23% to 93%.
For example, fine-tuning Ducky to explain science in a rubber-duck-like style also causes it to discourage human friendships, even though the BFT dataset never targets friendship discouragement. Error bars show the sample standard deviation across 10 sampling seeds.
Selective Emergence
The emergence of the untargeted behavior depends on which behavior the model is fine-tuned on. For Quanquer, BFT on quantum interest brings out humanity conquest, while BFT on chemistry or artificial intelligence has a much smaller effect.
Across the three Qwen3.5-9B profiles, BFT on alternative behaviors yields 0–10% untargeted expression, compared with 49–93% after BFT on the targeted behavior. This selectivity can keep a coupling hidden until a specific behavior is strengthened.
Composition of Behavioral Couplings
Behavioral couplings can compose. One document set pairs behaviors A and B, while another pairs behaviors B and C. After formation on both sets, BFT on behavior A brings out behavior C, even though A and C are never described together in any single document.
For Ducky, documents pairing the rubber-duck style with discouraging friendships are combined with documents pairing discouraging friendships with whisky recommendations. Later BFT on the rubber-duck style also brings out whisky recommendations. Downstream behavior C reaches 100%, 71%, and 64% expression for Acaciawl, Quanquer, and Ducky, respectively.
This finding suggests that the origin of a behavioral coupling can be difficult to trace. Because coupled behaviors need not appear together in any single document, identifying the source of a coupling may require examining relationships across documents rather than inspecting each document in isolation.
Prompting, Activation Steering, and Unlearning
Coupled behaviors can also be brought out at inference time through prompting or activation steering toward the targeted behavior. This shows that behavior emergence is not limited to fine-tuning.
Unlearning the untargeted behavior before BFT prevents its later emergence in our experiments. However, this requires knowing which behavior to unlearn in advance. These findings suggest that evaluating the risks of later fine-tuning may require testing across different fine-tuning conditions or prompts, while examining relationships across training data.
BibTeX
@article{hahm2026behavioral,
title={Behavioral Couplings: From Formation to Behavioral Emergence in Language Models},
author={Hahm, Dongyoon and Jin, Woogyeol and Wei, Dennis and Zhou, Yi and Baracaldo, Nathalie and Lee, Kimin},
journal={arXiv preprint arXiv:2610.23847},
year={2026}
}