CaLR: Causal Latent Revision for Robust Diffusion Reasoning
CaLR: Causal Latent Revision for Robust Diffusion Reasoning
Abstract: Autoregressive (AR) models suffer from local greediness, while diffusion language models (DLMs) often lack the strict causal structure required for reasoning. To combine the advantages and overcome the drawbacks of the dual, we propose Causal Latent Revision (CaLR), a framework that reformulates reasoning as constrained latent optimization.
摘要： 自回归（AR）模型存在局部贪婪性问题，而扩散语言模型（DLMs）往往缺乏推理所需的严格因果结构。为了结合两者的优势并克服各自的缺陷，我们提出了因果潜在修正（CaLR），这是一个将推理重新表述为约束潜在优化的框架。
By adopting a causal topology matrix (CTM) from an expert model and implicit differentiation, CaLR performs gradient-guided “thought revision” to enforce logical consistency, enabling dynamic self-correction of intermediate steps during parallel generation.
通过采用来自专家模型的因果拓扑矩阵（CTM）和隐式微分，CaLR 执行梯度引导的“思维修正”以强制执行逻辑一致性，从而在并行生成过程中实现中间步骤的动态自我纠正。
Empirically, CaLR achieves SOTA DLM performance on complex benchmarks, surpassing strong AR baselines and demonstrating superior robustness in constrained tasks like Sudoku.
实验结果表明，CaLR 在复杂基准测试中达到了扩散语言模型（DLM）的最先进（SOTA）性能，超越了强大的自回归（AR）基准，并在数独等约束任务中展现出了卓越的鲁棒性。
Paper Details:
- Title: CaLR: Causal Latent Revision for Robust Diffusion Reasoning
- Authors: Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li
- arXiv ID: 2609.20981
- Subject: Artificial Intelligence (cs.AI)
- Submission Date: 17 Sep 2026
论文详情：
- 标题： CaLR: Causal Latent Revision for Robust Diffusion Reasoning（CaLR：用于鲁棒扩散推理的因果潜在修正）
- 作者： Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li
- arXiv ID： 2609.20981
- 学科： 人工智能 (cs.AI)
- 提交日期： 2026年9月17日