Research map / Reinforcement learning / RL from feedback
Adaptive Code Watermarking Through Reinforcement Learning
Zhimeng Guo, Huaisheng Zhu, Siyuan Xu, Hangfan Zhang, Teng Xiao, Minhao Cheng
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 process-based advantage A2 (token-level reward) (ablation) | CodeTracer · Reward ablation | AUROC: 75.11 -> 82.95 | Improved |
| add process-based advantage A2 (token-level reward) (ablation) | CodeTracer · Reward ablation | TPR: 30.29 -> 46.34 | Improved |
| add process-based advantage A2 (token-level reward) (ablation) | CodeTracer · Reward ablation | Pass@1: 61.15 -> 60.82 | No clear effect |
| add outcome-based advantage A1 (ablation) | CodeTracer · Reward ablation | AUROC: 79.52 -> 82.95 | Improved |
| add outcome-based advantage A1 (ablation) | CodeTracer · Reward ablation | TPR: 34.91 -> 46.34 | Improved |
| add outcome-based advantage A1 (ablation) | CodeTracer · Reward ablation | Pass@1: 60.34 -> 60.82 | Improved |
| add supervised fine-tuning initialization before RL (CodeTracer-2 vs RL only CodeTracer-1) (ablation) | CodeTracer · HumanEval | Pass@1 / AUROC / TPR: Pass@1 62.65 -> 60.82; AUROC 77.71 -> 82.95; TPR 32.32 -> 46.34 | Mixed |
| apply watermark model trained on 1.5B directly to 8B LLM (vs WLLM) | CodeTracer · OpenCoder-8B-Instruct | AUROC: 65.90 -> 78.69 | Improved |
| CodeTracer vs SWEET under DIPPER paraphrasing attack | CodeTracer · HumanEval (DIPPER attack) | AUROC: 53.05 -> 58.42 | Improved |
| CodeTracer vs SWEET on Java zero-shot transfer | CodeTracer · HumanEvalPack Java | Pass@1: 25.22 -> 31.05 | Improved |
| CodeTracer vs SWEET on C++ zero-shot transfer | CodeTracer · HumanEvalPack C++ | Pass@1: 44.85 -> 44.39 | No clear effect |
Closest papers by meaning
- Learning to Follow In-Context Watermark Instructions via Self-Distillation (NeurIPS 2026 · similarity 0.86)
- StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning (ICML 2026 · similarity 0.85)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.85)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.85)
- No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices (NeurIPS 2024 · similarity 0.84)
- Watermark Stealing in Large Language Models (ICML 2024 · similarity 0.84)
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.
