CaLR：面向鲁棒扩散推理的因果潜在修正
原标题：CaLR: Causal Latent Revision for Robust Diffusion Reasoning
AI 摘要
该论文提出 CaLR（因果潜在修正）框架，将扩散语言模型（DLM）的推理过程重构为受约束的潜在空间优化。它通过专家模型提取因果拓扑矩阵（CTM），并利用隐式微分进行梯度引导的“思维修正”，从而在并行生成中实现中间步骤的动态自我纠错。实验表明，CaLR 在 GSM8K 和 Sudoku 等推理基准上取得 SOTA 表现，超越强自回归（AR）基线和标准 DLM 基线，并在严格约束任务中展现出更强的鲁棒性。
正文节选
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. By adopting a causal topology matrix (CTM) from an expert model and implicit differentia