https://chen-yingfa.github.io/categories/paper-note/Recent content in Paper Note on Yingfa Chen 陈英发Hugo -- 0.146.6en-usSat, 07 Oct 2023 17:55:33 +0000https://chen-yingfa.github.io/research_posts/2023-actadd/Sat, 07 Oct 2023 17:55:33 +0000https://chen-yingfa.github.io/research_posts/2023-actadd/<p><a href="https://arxiv.org/abs/2308.10248">Paper</a></p> <p>TLDR: Propose <strong>ActAdd</strong>, a method for controlling model behavior during inference by modifying activations with a bias term that is learned from a pair of prompt.</p> <p>Summary:</p> <ul> <li>Propose <strong>ActAdd</strong>, a method for controlling model behavior by modifying activations at inference time.</li> <li>Steering vectors are computed by taking the activation differences that result from pairs of prompts. The vectors are added as bias during inference.</li> <li>ActAdd provides control over high-level properties of the output, and preserves off-target model performance, and requires little computational and implementational costs.</li> </ul> <!-- more --> <blockquote> <p>The recently popular <a href="https://arxiv.org/abs/2310.01405">representation engineering paper</a> (RepE) seems to be largely inspired by this work.</p>