Research map / Large language models / Reasoning and chain of thought
HARDTESTGEN: A High-Quality RL Verifier Generation Pipeline for LLM Algorithimic Coding
Zhongmou He, Yee Man Choi, Kexun Zhang, Ivan Bercovich, Jiabao Ji, Junting Zhou, Dejia Xu, Aidan Zhang, Yixiao Zeng, Lei Li
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HardTestsGen, a pipeline for synthesizing hard test cases (verifiers) for LLM algorithmic coding RL reward signals.
Problem
Reliable verifiers for code generation are hard to build because wrong solutions may only be caught by carefully crafted edge cases that are difficult to synthesize automatically.
Approach
An automated pipeline (HardTestsGen) that synthesizes difficult test cases to serve as high-quality verifiers for RL on algorithmic coding, aimed at detecting well-disguised wrong programs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HardTests tests instead of TACO tests | verifier for Qwen2.5-Coder-7B programs · Codeforces/AtCoder 1253 problems | precision: 61 -> 79.21 | Improved |
| use HardTests tests instead of TACO tests, hardest problems | verifier · difficulty 4+, Qwen2.5-Coder-7B | precision: 17.83 -> 55.88 | Improved |
| add RPGen/SPGen programmed inputs to LLMGen (ablation) | HardTests tests · Qwen2.5-Coder-7B programs | precision: 42.42 -> 64.67 | Improved |
| add HackGen hacking inputs on top of LLMGen+RPGen/SPGen (ablation) | HardTests tests · Qwen2.5-Coder-7B programs | precision: 64.67 -> 79.21 | Improved |
| add HackGen hacking inputs (ablation) | HardTests tests · Qwen2.5-Coder-7B programs | recall: 92.01 -> 91.69 | No clear effect |
| use HardTests tests instead of CodeContests tests | verifier · human submissions | recall: 86.26 -> 93.03 | Improved |
| fine-tune on HardTests-verified correct trajectories (good 5k) vs bad 5k (ablation) | Qwen3-4B rejection sampling · LiveCodeBench-105 | pass@10: 54.92 -> 60.00 | Improved |
| rejection sampling with HardTests-verified correct data | Qwen3-4B · LiveCodeBench-105 | pass@1: 38.48 -> 36.00 | Hurt |
| rejection sampling with random 5k | Qwen3-4B · LiveCodeBench-105 | pass@1: 38.48 -> 32.75 | Hurt |
| RL with HardTests verifier | Qwen3-4B · LiveCodeBench-105 | pass@10: 56.19 -> 64.76 | Improved |
| RL with TACO verifier | Qwen3-4B · LiveCodeBench-105 | pass@1: 38.48 -> 36.95 | Hurt |
Closest papers by meaning
- Rethinking Verification for LLM Code Generation: From Generation to Testing (NeurIPS 2025 · similarity 0.92)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.90)
- Learning to Generate Unit Test via Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.89)
- CURE: Co-Evolving Coders and Unit Testers via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.89)
- Generative Verifiers: Reward Modeling as Next-Token Prediction (ICLR 2025 · similarity 0.88)
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.
