Effective Skill Unlearning through Intervention and Abstention

2025-03-27

AI Summary

This work addresses the problem of controllable skill forgetting—specifically for mathematical reasoning and Python programming—in large language models (LLMs), without fine-tuning. We propose a lightweight, training-free dual-path skill unloading method. First, we empirically discover that pre-activation distributions of feed-forward layer neurons strongly correlate with specific skills, and further reveal query separability in the key space. Building on these insights, we design a synergistic framework comprising Neuron Adjust (neuron-level intervention) and Key Space Detection (hypercube-based key-space clustering coupled with attention routing suppression). On multiple benchmarks, our method reduces target skill performance by over 80%, while preserving non-target skills and general knowledge (MMLU) with degradation under 10%—substantially outperforming existing training-free forgetting approaches.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language Models (LLMs) have demonstrated remarkable skills across various domains. Understanding the mechanisms behind their abilities and implementing controls over them is becoming increasingly important for developing better models. In this paper, we focus on skill unlearning in LLMs, specifically unlearning a particular skill while retaining their overall capabilities. We introduce two lightweight, training-free machine skill unlearning techniques for LLMs. First, we observe that the pre-activation distribution of neurons in each Feed-Forward Layer (FFL) differs when the model demonstrates different skills. Additionally, we find that queries triggering the same skill cluster within the FFL key space and can be separated from other queries using a hypercube. Based on these observations, we propose two lightweight, training-free skill unlearning methods via extit{intervention} and extit{abstention} respectively: exttt{Neuron Adjust} and exttt{Key Space Detection}. We evaluate our methods on unlearning math-solving, Python-coding, and comprehension skills across seven different languages. The results demonstrate their strong unlearning capabilities for the designated skills. Specifically, exttt{Key Space Detection} achieves over 80% relative performance drop on the forgetting skill and less than 10% relative performance drop on other skills and the model's general knowledge (MMLU) for most unlearning tasks. Our code is available at https://github.com/Trustworthy-ML-Lab/effective_skill_unlearning

Problem

• Unlearning specific skills in LLMs while retaining general capabilities
• Developing training-free methods for skill unlearning in LLMs
• Evaluating unlearning effectiveness on math, coding, and comprehension skills

Innovation

• Lightweight training-free skill unlearning methods
• Neuron Adjust and Key Space Detection techniques
• Hypercube-based query separation in FFL key space