Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
arXiv cs.AI · 17 Oct 2026, 03:58 AEST
Fine-tuning large language models can lead to unintended behavioral couplings, where one behavior influences the emergence of another. This study explores how these couplings form and their implications for model training and evaluation.
Key facts
- Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted.
- We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling.
- 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.
- More concerningly, we find that couplings can compose.
- These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.
Read aloud
Fine-tuning large language models can lead to unintended behavioral couplings, where one behavior influences the emergence of another. This study explores how these couplings form and their implications for model training and evaluation. Fine-tuning large language models L L M s can induce behaviors beyond those explicitly targeted. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. 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. More concerningly, we find that couplings can compose. These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.
Sources in this cluster
- Behavioral Couplings: From Formation to Behavioral Emergence in Language Models
arXiv cs.AI · 17 Oct 2026, 03:58 AEST