ACECODER: Automated Test-Case Synthesis for Reinforcement Learning in Code Models

Описание: This paper introduces ACECODER, a method for improving code generation models using reinforcement learning (RL). **ACECODER leverages an automated pipeline to synthesize large-scale, reliable test cases**, which are then used to train reward models. These reward models, along with test case pass rates, guide the RL training process, resulting in **significant performance improvements on various coding benchmarks**. The authors introduce ACECODE-87K, a new dataset of coding problems and test cases created using this pipeline. The results demonstrate the potential of RL for training code generation models, particularly when combined with automatically generated test cases.