Effective Skill Unlearning through Intervention and Abstention

In this work, we first observe that the pre-activation distribution of neurons in LLM's 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 intervention and abstention:

• Neuron Adjust unlearns one skill by shifting selected feed-forward layer neurons' pre-activation value from the forgetting skill distribution to the retaining skill distribution.
• Key Space Detection unlearns one skill by abstaining the model's output when the model's inference-time key vector falls in the forgetting skill hypercube.

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, 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 for most unlearning tasks.

Past works have shown that neurons represent different concepts, both in vision models and language models [1][2]. Pruning neurons (setting their values to 0) that represent unwanted concepts or skills could change the model's behavior [3][4]. However, recent works also show that neurons exhibit multisemanticity [5], with some being expressible as a linear combination of concepts [6], which means pruning those neurons could be harmful to the model's overall capabilities.

In this work, we observe that neurons exhibit different value distributions when the model demonstrates various skills. Additionally, we find that key vectors in the model's feed-forward layers (FFL) tend to cluster when the model performs different skills, allowing them to be bounded within a hypercube. Building on these observations, we propose two machine skill unlearning methods: Neuron Adjust and Key Space Detection, which achieve better performance in unlearning specific skills while preserving the model's overall capabilities.

Neuron Adjust consists of two steps:

1. Distribution Modeling: We probe the model with the forgetting and retaining datasets, record each token's preactivation value on each neuron, and model the whole value distribution using Gaussian distribution.
2. Inference-time Neuron Adjust: During inference time, we choose to adjust the top k% of the neurons with the most distributional difference. If we detect the neuron's preactivation value is more likely to be sampled from the forgetting distribution, we probablistically shift the neuron's value to the retaining distribution.

Figure 2. The overview of Neuron Adjust method.

            
@article{Li2025effective,
    title={Effective Skill Unlearning through Intervention and Abstention},
    author={Li, Yongce and Sun, Chung-En and Weng, Tsui-Wei},
    journal={NAACL},
    year={2025},
}