Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning
Authors: Yinjie Wang, Ling Yang, Ye Tian, Ke Shen, Mengdi Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive evaluations on five benchmarks and demonstrate that our CURE models, derived from base models of varying sizes, excel in both code generation and unit test generation. |
| Researcher Affiliation | Collaboration | Yinjie Wang1 , Ling Yang2 , Ye Tian3, Ke Shen4, Mengdi Wang2 1University of Chicago, 2Princeton University, 3Peking University, 4Byte Dance |
| Pseudocode | Yes | C.1 Detailed Algorithm We have our detailed CURE optimization pipeline as follows. The training data we use is our training split of Code Contests. We set n = m = 32, η = 1e 6 and β = 0.01. Algorithm 1 CURE |
| Open Source Code | Yes | Seed Project: https://github.com/Gen-Verse/CURE |
| Open Datasets | Yes | Datasets We select five widely used coding datasets for our comprehensive evaluation: Live Bench [45], MBPP [2], Live Code Bench [19], Code Contests [23], and Code Forces [31]. |
| Dataset Splits | Yes | For Code Contests, we extract tasks with difficulty level 2, and randomly split them into a training set of 4.5k examples and an evaluation set of 200 examples. |
| Hardware Specification | Yes | We train these models using 8 A100 GPUs. |
| Software Dependencies | No | The paper mentions models like Qwen2.5-7B and 14B, and Qwen3-4B, and a system called v LLM [20]. It also mentions Python for scripting. However, specific version numbers for these or other software libraries (e.g., PyTorch, TensorFlow, CUDA) are not provided in the extracted text. |
| Experiment Setup | Yes | At each sampling step during reinforcement learning, we generate 32 rollouts for unit tests and 32 for code using v LLM [20], with a temperature of 1.0, top-p of 0.95, and top-k of 40. For optimization, we set the learning rate to 1e-6 and the KL coefficient to 0.01. Specifically, for the long-Co T model, we use a lower temperature of 0.8 and apply a response-length-guided transformation to the unit test reward to improve post-training inference efficiency. |