New 'attention manifolds' offer steerable control over LLM outputs

Researchers have developed a novel technique called "attention manifolds" to steer and modify the behavior of large language models. These manifolds, implemented as learned B-spline surfaces, modulate value vectors based on query-key interactions, offering a more nuanced control than standard attention mechanisms. When applied to LLaMA models, this method improved perplexity on WikiText-2 and significantly altered outputs for a majority of diverse prompts, correcting factual errors, increasing precision, and adding specificity. The learned surfaces are also editable, allowing for direct model steering and the creation of "attention walls" to block specific value dimensions, which could have implications for AI safety. AI

IMPACT Enables more precise control and editing of LLM outputs, potentially improving factual accuracy and safety.

RANK_REASON The cluster contains an academic paper detailing a new method for controlling LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• LLaMA 3B-Instruct
• ScienceCast
• St. Augustine
• WikiText-2

• paper
• model release

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