arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.19871cs.CLcs.AI
收藏

基于流匹配的理由分布蒸馏用于分布连续隐式推理

Rationale Distribution Distillation via Flow Matching for Distributional Continuous Latent Reasoning

Korea Advanced Institute of Science and Technology(韩国科学技术院)

Mingyeong Kim, Gyusang Cho, Juho Lee, Giung Nam

AI总结:

该研究针对连续隐式推理的点值方法限制,提出理由分布蒸馏(RDD)方法,通过流匹配结合查询内混合提升答案覆盖,在高级数学任务上优于基线。

摘要:

从语言模型中采样多个思维链理由可发现单次尝试遗漏的正确解决方案,但这些理由通常冗长,导致重复自回归生成成本高昂。连续隐式推理提供了紧凑替代方案,但点值方法基于单个理由条件化重复答案采样,会限制对多样路径和答案覆盖的探索。因此,我们提出理由分布蒸馏(RDD),通过从多个文本推理路径中蒸馏,在固定长度隐式序列上训练条件流。此外,我们通过查询内混合扩展蒸馏目标,对同一查询的完整隐式理由进行插值。该流非自回归生成整个隐式序列,冻结的语言模型从采样的理由上下文自回归解码答案。实验表明,所得分布推理器比评估的点值隐式推理基线实现更高的答案覆盖,尤其在高级数学任务上表现突出。

英文摘要:

Sampling multiple chain-of-thought rationales from a language model can uncover correct solutions missed by a single attempt. However, these rationales are typically lengthy, making repeated autoregressive generation costly. Continuous latent reasoning offers a compact alternative, but point-valued methods condition repeated answer sampling on a single rationale, which can limit exploration of diverse paths and answer coverage. We thus propose rationale distribution distillation (RDD), training a conditional flow over fixed-length latent sequences by distilling from multiple textual reasoning paths. Furthermore, we expand the distillation targets through within-query mixup, interpolating between complete latent rationales for the same query. The flow nonautoregressively generates the entire latent sequence, and a frozen language model autoregressively decodes answers from sampled rationale contexts. Empirically, the resulting distributional reasoner achieves higher answer coverage than the evaluated point-valued latent reasoning baselines, particularly on advanced mathematical tasks.

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。

↑