Research map / Large language models / Instruction and fine-tuning
ATGen: Adversarial Reinforcement Learning for Test Case Generation
Qingyao Li, Xinyi Dai, Weiwen Liu, Xiangyang Li, Yasheng Wang, Ruiming Tang, Yong Yu, Weinan Zhang
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 |
|---|---|---|---|
| train test generator with GRPO RL instead of SFT (UTGen) | Qwen2.5-7B-Instruct · APPS+Codeforces subset | attack rate (%): 16.24 -> 34.02 | Improved |
| add adversarial code generation (Adaptive) (ablation) | ATGen 7B · APPS+Codeforces subset | attack rate (%): 34.02 -> 36.99 | Improved |
| add adversarial code generation (Unconditional) (ablation) | ATGen 3B · APPS+Codeforces subset | attack rate (%): 29.96 -> 32.38 | Improved |
| add adversarial code generation (Unconditional) (ablation) | ATGen 7B · APPS+Codeforces subset | attack rate (%): 34.02 -> 34.57 | No clear effect |
| add adversarial code generation (Adaptive) (ablation) | ATGen 7B · APPS+Codeforces subset (Hard) | attack rate (%): 18.42 -> 21.05 | Improved |
| add adversarial code generation (Adaptive) (ablation) | ATGen 7B · APPS+Codeforces subset | IO accuracy (%): 71.56 -> 74.42 | Improved |
| reward only the final attack instead of IO acc plus input attack (ablation) | ATGen (w/o Adver) 3B · APPS+Codeforces subset | IO accuracy (%): 44.67 -> 67.72 | Improved |
| vary reward composition (ablation) | ATGen (w/o Adver) 3B · APPS+Codeforces subset | attack rate (%): 30.07 / 29.74 / 30.29 | No clear effect |
| increase sampling retries 10 -> 30 (ablation) | ATGen 7B · APPS+Codeforces subset | IO accuracy (%): 74.09 -> 70.69 | Hurt |
| increase sampling retries 10 -> 30 (ablation) | ATGen 7B · APPS+Codeforces subset | input attack rate (%): 47.96 -> 49.94 | Improved |
| use sampling-based instead of instruction-based adversarial code (ablation) | ATGen (Adaptive) 7B · APPS+Codeforces subset | attack rate (%): 32.71 -> 36.22 | Improved |
| use sampling-based instead of instruction-based adversarial code (ablation) | ATGen (Adaptive) 7B · APPS+Codeforces subset | IO accuracy (%): 76.28 -> 74.09 | Hurt |
Closest papers by meaning
- Learning to Generate Unit Test via Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.92)
- Rethinking Verification for LLM Code Generation: From Generation to Testing (NeurIPS 2025 · similarity 0.89)
- CURE: Co-Evolving Coders and Unit Testers via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models (ICML 2026 · similarity 0.86)
- Teaching Language Models to Critique via Reinforcement Learning (ICML 2025 · similarity 0.86)
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.86)
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.
