Lishui Fan

Lishui Fan, Yu Zhang, Mouxiang Chen et al.

Reinforcement learning (RL) has significantly advanced code generation for large language models (LLMs). However, current paradigms rely on outcome-based rewards from test cases, neglecting the quality of the intermediate reasoning process. While supervising the reasoning process directly is a promising direction, it is highly susceptible to reward hacking, where the policy model learns to exploit the reasoning reward signal without improving final outcomes. To address this, we introduce a unified framework that can effectively incorporate the quality of the reasoning process during RL. First, to enable reasoning evaluation, we develop LCB-RB, a benchmark comprising preference pairs of superior and inferior reasoning processes. Second, to accurately score reasoning quality, we introduce an Optimized-Degraded based (OD-based) method for reward model training. This method generates high-quality preference pairs by systematically optimizing and degrading initial reasoning paths along curated dimensions of reasoning quality, such as factual accuracy, logical rigor, and coherence. A 7B parameter reward model with this method achieves state-of-the-art (SOTA) performance on LCB-RB and generalizes well to other benchmarks. Finally, we introduce Posterior-GRPO (P-GRPO), a novel RL method that conditions process-based rewards on task success. By selectively applying rewards to the reasoning processes of only successful outcomes, P-GRPO effectively mitigates reward hacking and aligns the model's internal reasoning with final code correctness. A 7B parameter model with P-GRPO achieves superior performance across diverse code generation tasks, outperforming outcome-only baselines by 4.5%, achieving comparable performance to GPT-4-Turbo. We further demonstrate the generalizability of our approach by extending it to mathematical tasks. Our models, dataset, and code are publicly available.

Lishui Fan, Mouxiang Chen, Tingwei Zhu et al.

Code generation is important in software engineering, and Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful paradigm to improve it through execution-based feedback. However, most RLVR pipelines rely on human-curated tests, making progress bottlenecked by scarce and costly supervision. Existing work tried to use self-generated tests to ground rewards, but the lack of discriminative tests constrains the effect due to the sub-optimal performance of the model on test generation. We aim to improve code generation without ground-truth supervision by co-evolving code and test generation, so that their interactions yield progressively more informative supervision. To this end, we present ZeroCoder, a fully label-free co-evolutionary framework that jointly trains a Coder and a Tester using execution feedback from self-generated code-test interactions. For each problem, ZeroCoder executes sampled solutions against sampled tests to form a passing matrix, identifies a consensus subset of likely-correct solutions and consistent tests via a pluggable selection algorithm, and derives role-specific rewards. To ensure reward quality, ZeroCoder filters low-information instances via rank-based pre-filtering and trains the Tester with a curriculum balancing validity and mutation-driven discriminativeness. We further identify selector drift, the progressive miscalibration of fixed selection rules during co-evolution, and introduce DyB4, a Bayesian selector that uses as few as 10 labeled instances to recalibrate its priors dynamically. Across three models and six benchmarks, ZeroCoder consistently improves code generation and test generation. In the fully label-free setting, it improves code generation by up to 14.5% over the base model on Qwen2.5-Coder-7B-Instruct. With DyB4, the gain reaches 21.6%, while test generation improves by 24.3%, approaching oracle-supervised performance.

Lishui Fan, Mouxiang Chen, Zhongxin Liu

Large language models (LLMs) have achieved impressive performance in code generation. However, due to the long-tail distribution of LLMs' training data, low-frequency terms are typically underrepresented in the training process. Consequently, LLMs often misunderstand or overlook problem-specific, low-frequency keywords during code generation, compromising the accuracy of the generated code. To address this, we propose a novel technique named SEK(\textbf{S}elf-\textbf{E}xplained \textbf{K}eywords), which empowers an LLM for better code generation by extracting and explaining the key terms in the problem description with the LLM itself and ranking them based on frequency. Comprehensive experiments across three benchmarks, i.e., HumanEval(+), MBPP(+), and APPS, with five representative LLMs, show that SEK can significantly improve LLMs in code generation, yielding substantial and consistent gains. For instance, SEK improves the Pass@1 of DeepSeek-Coder-V2-Instruct from 85.4\% to 93.3\% on the Humaneval benchmark. Further analysis confirms that SEK enables the LLMs to shift their attention from low-frequency keywords to their corresponding high-frequency counterparts.

Hanzhen Lu, Lishui Fan, Jiachi Chen et al.

Line-level code completion requires a critical balance between high accuracy and low latency. Existing methods suffer from a trade-off: large language models (LLMs) provide high-quality suggestions but incur high latency, while small language models (SLMs) are fast but often suboptimal. We propose MCCom (Model-Cascading-based code Completion), a framework that cascades a local SLM with a cloud-based LLM. To achieve effective cascading, MCCom leverages user actions as a novel signal to trigger the LLM only when the SLM fails, significantly reducing cloud computation costs. Furthermore, we introduce a two-stage speculative decoding strategy and an iterative retrieval mechanism to enhance collaboration between the models. We also train a 121M-parameter lightweight model, which achieves 73.8% of the performance of a 7B state-of-the-art model. Evaluated on RepoEval and a new real-world benchmark StmtEval, MCCom reduces inference latency by up to 47.9% and LLM usage by 46.3%, while improving the LLM's exact match rate by 8.9% through effective collaboration.