Activation Addition: Steering Language Models Without Optimization
Authors:
Alexander Matt Turner,
Lisa Thiergart,
Gavin Leech,
David Udell,
Juan J. Vazquez,
Ulisse Mini,
Monte MacDiarmid
Abstract:
Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implici…
▽ More
Reliably controlling the behavior of large language models is a pressing open problem. Existing methods include supervised finetuning, reinforcement learning from human feedback, prompt engineering and guided decoding. We instead investigate activation engineering: modifying activations at inference-time to predictably alter model behavior. We bias the forward pass with a 'steering vector' implicitly specified through natural language. Past work learned these steering vectors; our Activation Addition (ActAdd) method instead computes them by taking activation differences resulting from pairs of prompts. We demonstrate ActAdd on a range of LLMs (LLaMA-3, OPT, GPT-2, and GPT-J), obtaining SOTA on detoxification and negative-to-positive sentiment control. Our approach yields inference-time control over high-level properties of output like topic and sentiment while preserving performance on off-target tasks. ActAdd takes far less compute and implementation effort than finetuning or RLHF, allows users control through natural language, and its computational overhead (as a fraction of inference time) appears stable or improving over increasing model size.
△ Less
Submitted 4 June, 2024; v1 submitted 20 August, 2023;
originally announced August 2023.