AI Paper Atlas

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

AAAI 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace sample-level unlearning (GA, NPO, FLAT, DPO) with token-level PRODCodeLlama-7B · Copyrighted code unlearningPDR (%): 15.3 (best baseline GA) -> 41.8Improved
replace baselines with PRODCodeLlama-7B · Insecure code unlearningPDR (%): 24.5 (best baseline NPO) -> 70.4Improved
replace baselines with PRODCodeLlama-7B · Deprecated API unlearningPDR (%): 51.6 (best baseline NPO) -> 58.0Improved
use cross-entropy instead of KL or JS divergence as PROD loss (ablation)PROD · Copyrighted code unlearningforget quality vs model utilityImproved
add moderate noise elimination (nucleus p = 0.8) versus none (ablation)PROD · Copyrighted code unlearningforget quality vs model utilityImproved
set alpha-suppression to 0 (no extra negative pressure) (ablation)PROD · Copyrighted code unlearningforget quality vs model utilityImproved
use positive alpha values (ablation)PROD · Copyrighted code unlearningforget quality vs model utilityMixed
use small negative alpha values (ablation)PROD · Copyrighted code unlearningforgetting speedMixed
apply GA, NPO, DPO, FLAT to code at best forget qualityCodeLlama-7B · HumanEval after copyright unlearningpass rate: near-zeroHurt
use PROD under prefix injection attackCodeLlama-7B · Copyrighted code prefix injection attacksimilarity to forgotten code: stays below 0.05 vs others exceed 0.3Improved
apply PROD to Qwen2.5-Coder-7B, Deepseek-coder-6.7B, Starcoder-7BPROD · Copyrighted code unlearningforget quality and utilityImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy