LLM Analogy Vector Unlocked by Instruction Exposure

Researchers have identified an "Analogy Vector" within large language models that represents a latent rhetorical strategy. This vector, accessible geometrically in the model's residual stream, can be used to steer the model's explanatory style. The study found that while the analogy direction exists in all base models, its behavioral accessibility scales with the amount of instruction exposure during training, challenging the traditional base vs. instruct model dichotomy. AI

IMPACT Reveals how instruction exposure shapes LLM rhetorical capabilities, potentially guiding future model training and interpretability methods.

RANK_REASON Research paper published on arXiv detailing a new interpretability technique for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Analogy Vector
• arXiv
• Gemma-2
• Gemma-2-9B-IT
• Hugging Face
• Llama-3.1
• Qwen-2.5

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →