Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model

Jinwei Hu, Zhenglin Huang, Xiangyu Yin, Wenjie Ruan, Guangliang Cheng, Yi Dong, Xiaowei Huang

TL;DR

FALCON addresses the challenge of safely removing specific knowledge from large language models by introducing a principled, fine-grained unlearning framework. It combines information-theoretic guidance to locate minimally entangled layers, contrastive representation unalignment via Principal Offset Vectors derived from SVD, and gradient projection to orthogonalize forgetting and retention updates, with a second-order optimizer for stability. Across harmful knowledge, copyrighted content, and entity unlearning tasks, FALCON achieves superior unlearning effectiveness while preserving general capabilities and showing resistance to knowledge recovery, demonstrating practical viability for responsible AI. While evaluated on text-based LLMs and smaller models, the approach offers a scalable path toward precise, interpretable, and robust unlearning in real-world settings with regulatory and safety implications.

Abstract

Large language models have been widely applied, but can inadvertently encode sensitive or harmful information, raising significant safety concerns. Machine unlearning has emerged to alleviate this concern; however, existing training-time unlearning approaches, relying on coarse-grained loss combinations, have limitations in precisely separating knowledge and balancing removal effectiveness with model utility. In contrast, we propose Fine-grained Activation manipuLation by Contrastive Orthogonal uNalignment (FALCON), a novel representation-guided unlearning approach that leverages information-theoretic guidance for efficient parameter selection, employs contrastive mechanisms to enhance representation separation, and projects conflict gradients onto orthogonal subspaces to resolve conflicts between forgetting and retention objectives. Extensive experiments demonstrate that FALCON achieves superior unlearning effectiveness while maintaining model utility, exhibiting robust resistance against knowledge recovery attempts.

FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model

TL;DR

FALCON addresses the challenge of safely removing specific knowledge from large language models by introducing a principled, fine-grained unlearning framework. It combines information-theoretic guidance to locate minimally entangled layers, contrastive representation unalignment via Principal Offset Vectors derived from SVD, and gradient projection to orthogonalize forgetting and retention updates, with a second-order optimizer for stability. Across harmful knowledge, copyrighted content, and entity unlearning tasks, FALCON achieves superior unlearning effectiveness while preserving general capabilities and showing resistance to knowledge recovery, demonstrating practical viability for responsible AI. While evaluated on text-based LLMs and smaller models, the approach offers a scalable path toward precise, interpretable, and robust unlearning in real-world settings with regulatory and safety implications.

Abstract

Large language models have been widely applied, but can inadvertently encode sensitive or harmful information, raising significant safety concerns. Machine unlearning has emerged to alleviate this concern; however, existing training-time unlearning approaches, relying on coarse-grained loss combinations, have limitations in precisely separating knowledge and balancing removal effectiveness with model utility. In contrast, we propose Fine-grained Activation manipuLation by Contrastive Orthogonal uNalignment (FALCON), a novel representation-guided unlearning approach that leverages information-theoretic guidance for efficient parameter selection, employs contrastive mechanisms to enhance representation separation, and projects conflict gradients onto orthogonal subspaces to resolve conflicts between forgetting and retention objectives. Extensive experiments demonstrate that FALCON achieves superior unlearning effectiveness while maintaining model utility, exhibiting robust resistance against knowledge recovery attempts.

Paper Structure

This paper contains 56 sections, 16 equations, 4 figures, 11 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. LLM Unlearning
  4. Problem Formulation
  5. Problem Setup
  6. LLM unlearning with MI Guidance
  7. Methodology
  8. Information-Theoretic Guidance for Unlearning
  9. Contrastive Orthogonal Unalignment
  10. Contrastive Representation Unlearning
  11. Orthogonalizing Gradient for Conflict Resolution
  12. Experiments
  13. Harmful Knowledge Unlearning
  14. Mutual Information for Parameter Selection
  15. Visualization of MI for LLMs
  16. ...and 41 more sections

Figures (4)

  • Figure 1: Schematic overview of FALCON. The pipeline comprises three stages: parameter selection based on mutual information (Step 1); contrastive orthogonal unalignment, which consists of contrastive mechanism on both forgetting and retention datasets (Step 2.1) and orthogonal gradient conflict resolution (Step 2.2); and model unlearning guided by these components (Step 3).
  • Figure 2: Heatmaps of MI across LLM layers show that lower MI values indicate layers better suited for unlearning, with early layers being more domain-specific and deeper layers more entangled.
  • Figure 3: Gradient conflicts across layers with minimum (blue) and maximum (orange) MI values computed during parameter selection in Mistral-7B.
  • Figure 4: Logit lens probing results on different components.