Large language models (LLMs) can be guided using compact task vectors, but these vectors are treated as black boxes with unknown internal structure. This makes them hard to understand or control precisely.
The researchers decomposed each task vector into two meaningful components: Task Recognition, which identifies the task from examples, and Task Learning, which learns input-output mappings. They then optimized the balance between these components to improve guidance.
The decomposed components produced more reliable performance improvements than arbitrary adjustments. This reveals that task vectors have interpretable internal structure, enabling better understanding and precise control of LLM behavior.