SynLat：句法对齐的文本-潜在思维链压缩
原标题：SYNLAT: Syntax-Aligned Text-Latent Compression for Chain-of-Thought Reasoning
AI 摘要
论文提出 SynLat，一种文本-潜在混合的思维链压缩框架，通过句法对齐单元（SAU）将压缩边界与句法结构对齐，避免 token 级或固定长度边界割裂短语、公式等连贯片段。方法用答案条件化的 Teacher 为单一压缩条件化 Student 构建渐进式 KEEP/LATENT 目标，推理时仅需问题和压缩级别即可生成混合推理。在 Qwen3-8B/14B 上，MEDIUM 级别提升 3.6/2.6 分，HIGH 级别提升 7.0/5.5 分，压缩越强优势越明显。
正文节选
SynLat: Syntax-Aligned Text–Latent Compression for Chain-of-Thought Reasoning Abstract Long chain-of-thought (CoT) traces impose substantial output-token costs. Under constrained budgets, compression must preserve answer-critical information, making boundary placement central. Token-level and fixed-length boundaries can fragment coherent spans such as phrases, formulas, and local derivations, whereas step-level boundaries can bind content requiring different compression actions. We introduce Syn