Steering LLM Rhetorical Style via Analogy Vector Extraction

Researchers identified an Analogy Vector within the residual stream of models like Llama-3.1 and Qwen-2.5. Linear intervention on this vector improves reasoning-by-analogy performance, increasing the harmonic mean from 3.22 to 3.94 for Gemma-2-9B-IT.