New framework enhances LLM activation steering with orthogonalized estimator

Researchers have introduced Neyman Orthogonal Activation Steering (NOAS), a novel framework for manipulating Large Language Model (LLM) activations. This approach disentangles the target estimand from the estimator, allowing for more flexible modeling, including continuous outcomes like log-probabilities. NOAS builds upon existing methods by framing them as plug-in estimators of the average gradient of an outcome regression and introduces a complementary estimator based on the Stein score. The resulting orthogonalized estimator offers double-robustness properties and is demonstrated on benchmark datasets. AI

IMPACT This new framework could enable more precise control over LLM outputs and improve the generation of synthetic data.

RANK_REASON The item is a research paper detailing a new method for LLM activation steering. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• activation steering
• binary contrastive label
• continuous outcomes
• Debiased Machine Learning for Estimating the Causal Effect of Urban Traffic on Pedestrian Crossing Behavior
• Flexible ML models
• gradient
• Gradient Functionalization of Various Quaternized Polyethylenimines on Microfluidic Chips for the Rapid Appraisal of Antibacterial Potencies
• LLM activations
• log-probabilities
• Neyman Orthogonal Activation Steering
• oracle ratings
• outcome regression
• Stein score
• synthetic data
• target concept

• paper
• model release

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