Research map / Large language models / Language, safety and interpretability
Autoregressive, Yet Revisable: In Decoding Revision for Secure Code Generation
Chengran Yang, zichao wei, Heminghao Deng, Jinfeng Jiang, Zhensu Sun, Ting Zhang, Tianyi Wu, Ming Wen, David Lo
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 |
|---|---|---|---|
| add Stream of Revision in-decoding revision tokens | Qwen2.5-7B · CyberSecEval 2 C | SPR@1: 64.36 -> 73.76 | Improved |
| add Stream of Revision | Qwen2.5-3B · CyberSecEval 2 C | SPR@1: 62.38 -> 75.25 | Improved |
| train with revision sequences instead of patched functions only (ablation) | Qwen2.5-7B · CyberSecEval 2 C | SPR: 66.48% -> 73.76 | Improved |
| single-pass revision instead of post-hoc agent with localization | Qwen2.5-7B · CyberSecEval 2 | input tokens: 742.53 -> 113.10 | Improved |
| single-pass revision instead of post-hoc agent with localization | Qwen2.5-7B · CyberSecEval 2 C | SPR: 70.87 vs 73.76 | Mixed |
| use larger relaxed training set (over 8,000 positive samples) (ablation) | Qwen2.5-7B · CyberSecEval 2 | SPR and tokens: SPR 73.76% -> 74.22% (C); tokens 280.94 -> 321.74 | No clear effect |
| add Stream of Revision on top of ProSec | Prosec · CyberSecEval 2 C++ | SPR: 79.75% -> 83.90% | Improved |
| raise revision trigger logit bias from 1 to 10 (ablation) | Qwen2.5-7B · CyberSecEval 2 | SPR and tokens: +1.41% SPR, 1.54x tokens | Mixed |
| replace substring-constrained localization with unconstrained (ablation) | Stream of Revision · CyberSecEval 2 | hallucinated spans: 10.14% (36/355) hallucinated | Hurt |
| train on C/C++ only, test on other languages | Stream of Revision · CyberSecEval 2 OOD languages | SPR@1: 68.24 -> 75.65 (Qwen2.5-7B all-language avg) | Improved |
| add Stream of Revision | Qwen2.5-7B · HumanEval C/C++ | Pass@10: 81.98 -> 79.50 | No clear effect |
Closest papers by meaning
- Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs (NeurIPS 2025 · similarity 0.86)
- CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules (ICLR 2024 · similarity 0.86)
- BaxBench: Can LLMs Generate Correct and Secure Backends? (ICML 2025 · similarity 0.85)
- AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement (ICML 2025 · similarity 0.85)
- Code Repair with LLMs gives an Exploration-Exploitation Tradeoff (NeurIPS 2024 · similarity 0.85)
- Training Language Models to Generate Quality Code with Program Analysis Feedback (NeurIPS 2025 · similarity 0.85)
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.
