arXiv 2605.196652026-07-10cs.SEcs.AI

CriterAlign: 以标准为中心的推理对齐用于代码偏好判断

CriterAlign: Criterion-Centric Rationale Alignment for Code Preference Judging

KAUST（卡塔尔人工智能研究 institute）；ByteDance（字节跳动）

Zhenyu Li, Aleksandar Cvejic, Zehui Chen, Peter Wonka

AI总结：

本文提出CriterAlign，一种以标准为中心的推理对齐框架，通过直接的标准级 pairwise 判断、tie-driven 标准细化、swap-consistency 过滤和最终 pairwise 合成，改进了代码偏好判断的准确性，同时引入Human-Preference-Aligned Guidance (HPAG)来提升性能。

摘要：

成对的人类偏好预测是评估代码生成系统的核心，其中质量往往依赖于任务特定的权衡，而不仅仅是功能正确性。虽然基于评分表的LLM判断通过将评估分解为显式标准来提高可解释性，但大多数现有流程仍然是逐点的：它们独立评分每个响应，并通过比较聚合分数来推导偏好。我们证明这种设计与成对的代码偏好预测不匹配，并且可能在强单体判断下表现不佳。我们提出了CriterAlign，一种以标准为中心的框架，通过直接的标准级成对判断、tie驱动的标准细化、swap一致性过滤和最终成对合成，将基于评分表的判断适应于成对偏好评估。我们进一步引入Human-Preference-Aligned Guidance (HPAG)，通过从训练示例中提取人类偏好与单体判断预测之间的反复推理缺口进行离线合成，并注入到标准生成器、标准判断器和最终判断器中。在BigCodeReward上，CriterAlign将Qwen2.5-VL-32B单体判断的准确率从60.4%提升到66.3%，消融实验确认了成对标准设计和HPAG的贡献。

英文摘要：

Pairwise human preference prediction is central to evaluating code-generation systems, where quality often depends on task-specific trade-offs beyond functional correctness. While rubric-based LLM judges improve interpretability by decomposing evaluation into explicit criteria, most existing pipelines remain pointwise: they score each response independently and derive preferences by comparing aggregated scores. We show that this design is poorly matched to pairwise code preference prediction and can underperform a strong monolithic judge. We propose CriterAlign, a criterion-centric framework that adapts rubric-based judging to pairwise preference evaluation through direct criterion-level pairwise judgments, tie-driven criterion refinement, swap-consistency filtering, and final pairwise synthesis. We further introduce Human-Preference-Aligned Guidance (HPAG), synthesized offline from training examples by extracting recurring rationale gaps between human preferences and monolithic judge predictions, and injected into the criterion generator, criterion judge, and final judge. On BigCodeReward, CriterAlign improves a Qwen2.5-VL-32B monolithic judge from 60.4% to 66.3% accuracy, with ablations confirming the contributions of pairwise criterion design and HPAG.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。