RobustTests Improves LLM Code Generation with Better Test Cases

Reinforcement learning from verifiable rewards (RLVR) is crucial for improving Large Language Model (LLM) code generation. However, its effectiveness is often hampered by inadequate test cases, which can lead to false positives, reward hacking, and a decline in policy quality. Current automated test generation methods often fall short in comprehensiveness. The RobustTests framework addresses these limitations by introducing a novel faulty-code-driven test case synthesis strategy. It leverages "near correct" faulty code to guide the model in identifying subtle logical errors. Additionally, it employs validator agents with behavioral feature clustering to meticulously filter out invalid or redundant test cases, ensuring a high-quality test suite. To counter false negatives arising from LLM hallucination in synthetic tests, RobustTests incorporates a stepwise dense reward function based on pass rates. This provides fine-grained feedback, bolstering training robustness. Experiments show that fine-tuning Qwen3-32B with RobustTests yields a 3% performance gain on LiveCodeBench, confirming its efficacy in advancing LLM code generation.