Fast & Faithful Function Vectors
Researchers have developed a new method for creating function vectors (FVs), which are task representations used to guide large language models (LLMs) during in-context learning. Their approach involves using gradient-based attributions with Layer-wise Relevance Propagation (LRP) for more efficient and accurate head selection. Additionally, they found that applying FV steering in a distributed manner improves accuracy compared to simple aggregation. AI
IMPACT This research could lead to more efficient and accurate steering of LLMs for specific tasks, improving their performance in in-context learning scenarios.