https://lucasboscaini.com/tags/llm/Recent content in LLM on Lucas Boscaini.comHugoen-usFri, 25 Sep 2026 00:00:00 +0000https://lucasboscaini.com/posts/post-training-free-task-vectors/Fri, 25 Sep 2026 00:00:00 +0000https://lucasboscaini.com/posts/post-training-free-task-vectors/<h1 id="training-free-task-vectors-for-controlling-llm-behavior">Training-Free Task Vectors for Controlling LLM Behavior</h1> <p>I am excited to share our latest research, <strong>&ldquo;Training-Free Task Vectors for LLM Behavioral Control,&rdquo;</strong> which has been accepted at <strong>NeurIPS 2026</strong>!</p> <p>In this paper, co-authored with Gabriel J. Perin, Lucas Boscaini, André Araujo, and Nina S. T. Hirata, we tackle a fundamental challenge in post-training model editing: <em>How can we permanently steer LLM behaviors without the heavy computational cost of fine-tuning?</em></p> <hr> <h2 id="the-problem-the-high-cost-of-classical-task-vectors">The Problem: The High Cost of Classical Task Vectors</h2> <p>Task vectors have emerged as a powerful paradigm for model editing. Traditionally, a task vector is constructed by taking the parameter-level difference between a fine-tuned model and its base checkpoint ( $\Delta\theta = \theta_{\text{fine-tuned}} - \theta_{\text{base}}$ ).</p>