ALTER: Asymmetric LoRA for Token-Entropy-Guided Unlearning of LLMs

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• ALTER achieves over 95% forget quality on the TOFU, WMDP, and MUSE benchmarks.result 95%
• ALTER preserves over 90% of model utility while unlearning.result 95%
• ALTER utilizes an asymmetric LoRA architecture for LLM unlearning.method 95%
• ALTER decouples unlearning from the LLM's billion-scale parameters, leading to excellent efficiency.technical 90%
• The proposed method isolates and unlearns specific token knowledge by separating high and low entropy tokens.method 90%
• The complexity and overhead of integrating ALTER with existing pretrained models might be significant.limitation 80%
• Performance on real-world, unseen data requires thorough evaluation.limitation 80%
• ALTER can replace traditional unlearning methods that risk essential knowledge loss or require extensive retraining.technical 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card