Flow matching has shown strong promise as a continuous generative framework, but its standard formulation is not well-suited to reasoning problems. Conventional flow matching trains a model to transport noise directly toward a single final target. For tasks that require multi-step reasoning, this direct path is often too difficult: the final answer may be reachable only through a sequence of intermediate computations or transformations. Inspired by how autoregressive language models address this challenge through chain-of-thought supervision, we introduce ReasonFlow, a framework that brings the same scaffolding principle to flow matching. Instead of supervising the flow with only a final target, ReasonFlow constructs a sequence of CoT-induced anchor points derived from intermediate reasoning prefixes. During training, the clean target moves along a piecewise-linear path through these anchors before endpoint answer readout. This turns reasoning supervision from an output sequence into a trajectory-shaping signal for flow matching. Across arithmetic, chemistry, and logical-reasoning tasks, ReasonFlow substantially improves over no-CoT flow-matching baselines.

The past years have witnessed a proliferation of large language models (LLMs). Yet, reliable evaluation of LLMs is challenging due to the inaccuracy of standard metrics in human perception of text quality and the inefficiency in sampling informative test examples for human evaluation. This paper presents a sample-efficient human evaluation method for LLMs based on the principle of MAximum Discrepancy (MAD) competition. MAD automatically selects a small set of informative input instructions, each of which maximizes the discrepancy of two LLMs’ reponses, which are subsequently subject to three-alternative forced choice by human subjects. The pairwise comparison results of multiple LLMs are then aggregated into a global ranking using the Elo rating system. We compare eight representative LLMs in terms of four skills: knowledge understanding, mathematical reasoning, writing, and coding. Experimental results show that the proposed method reliably achieves the “golden” ranking of LLMs with a minimum set of input instructions, which in turn reveal their relative strengths and weaknesses, and offers valuable insights for further LLM advancement.