Research map / Large language models / Language, safety and interpretability
Large Language Model Unlearning for Source Code
Xue Jiang, Yihong Dong, Huangzhao Zhang, Tangxinyu Wang, Zheng Fang, Yingwei Ma, Rongyu Cao, Binhua Li, Zhi Jin, Wenpin Jiao, Yongbin Li, Ge Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace sample-level unlearning (GA, NPO, FLAT, DPO) with token-level PROD | CodeLlama-7B · Copyrighted code unlearning | PDR (%): 15.3 (best baseline GA) -> 41.8 | Improved |
| replace baselines with PROD | CodeLlama-7B · Insecure code unlearning | PDR (%): 24.5 (best baseline NPO) -> 70.4 | Improved |
| replace baselines with PROD | CodeLlama-7B · Deprecated API unlearning | PDR (%): 51.6 (best baseline NPO) -> 58.0 | Improved |
| use cross-entropy instead of KL or JS divergence as PROD loss (ablation) | PROD · Copyrighted code unlearning | forget quality vs model utility | Improved |
| add moderate noise elimination (nucleus p = 0.8) versus none (ablation) | PROD · Copyrighted code unlearning | forget quality vs model utility | Improved |
| set alpha-suppression to 0 (no extra negative pressure) (ablation) | PROD · Copyrighted code unlearning | forget quality vs model utility | Improved |
| use positive alpha values (ablation) | PROD · Copyrighted code unlearning | forget quality vs model utility | Mixed |
| use small negative alpha values (ablation) | PROD · Copyrighted code unlearning | forgetting speed | Mixed |
| apply GA, NPO, DPO, FLAT to code at best forget quality | CodeLlama-7B · HumanEval after copyright unlearning | pass rate: near-zero | Hurt |
| use PROD under prefix injection attack | CodeLlama-7B · Copyrighted code prefix injection attack | similarity to forgotten code: stays below 0.05 vs others exceed 0.3 | Improved |
| apply PROD to Qwen2.5-Coder-7B, Deepseek-coder-6.7B, Starcoder-7B | PROD · Copyrighted code unlearning | forget quality and utility | Improved |
Closest papers by meaning
- A Closer Look at Machine Unlearning for Large Language Models (ICLR 2025 · similarity 0.90)
- On Large Language Model Continual Unlearning (ICLR 2025 · similarity 0.90)
- ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding (ICLR 2025 · similarity 0.90)
- OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of Methods and Metrics (NeurIPS 2025 · similarity 0.89)
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code (ICLR 2025 · similarity 0.89)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
