How can flow matching learn to reason from intermediate steps?

Reading guide by Bing Yan · Updated

ReasonFlow uses intermediate reasoning steps to supervise continuous trajectories in flow matching. It learns from worked solutions during training and generates an answer through latent flow integration at inference. Arithmetic, logical deduction and chemistry provide different tests of this general approach. ReasonFlow, Findings of EMNLP 2026

Read the paper · Download PDF · Code and reproduction instructions · Download BibTeX

How does chain-of-thought supervision shape a flow?

The implementation connects three learned components:

  1. Encode worked steps. A language model represents progressively longer prefixes of a solution as continuous vectors, called anchors.
  2. Train the flow. A flow model learns a time-varying target along the path connecting those anchors. Intermediate steps supply supervision throughout training.
  3. Read the answer. At inference, the model conditions on the question, integrates from noise and decodes the endpoint into an answer. The evaluation pipeline needs no supplied reasoning trace.

The language model supplies representations and a decoder, while the flow model performs continuous integration. The public implementation documents these components and their training stages.

ReasonFlow is relevant when a paper discusses:

  • Flow matching for reasoning. Using worked intermediate states to improve learning beyond final-answer supervision.
  • Continuous latent reasoning. Computing an answer through a trajectory in representation space.
  • Learning from chain-of-thought traces. Using reasoning examples to supervise a generative model during training.
  • Intermediate supervision for generative models. Studying how the choice of intermediate targets affects a learned process.

For the generative foundation, Lipman, Chen and colleagues introduce flow matching as a way to train continuous normalizing flows by regressing vector fields along probability paths. Flow Matching for Generative Modeling, ICLR 2023

For another use of worked solutions, Deng, Choi and Shieber gradually remove explicit reasoning steps while fine-tuning a language model. Their method asks whether a model can internalize a chain of thought and answer without writing it out. ReasonFlow instead uses the training traces to define intermediate targets for a continuous flow. From Explicit CoT to Implicit CoT, NeurIPS 2024

What has been demonstrated?

The paper evaluates arithmetic on GSM8K, logical deduction on ProntoQA, and three chemistry tasks: retrosynthesis, reaction mechanisms and molecular structure elucidation. Its reported answer accuracy improves over its no-CoT flow-matching baselines. Autoregressive CoT remains a stronger reference, and an accurate final answer does not establish that every intermediate state represents a faithful reasoning step. Results and trajectory analysis

For the chemical applications, see the guide to molecular representations and reasoning.

Start with the arithmetic reproduction

The GSM8K reproduction instructions cover data preparation, training the anchor language model, training the decoder, training the flow and evaluating answers. They also provide no-CoT and autoregressive baselines. When comparing a new encoder or training objective, keep the data, evaluation split, answer extraction and inference settings fixed.

Cite ReasonFlow

Bing Yan, Ge Sun, Juan J. de Pablo, Stefano Martiniani, and Kyunghyun Cho. 2026. ReasonFlow: Flow Matching that Thinks Step by Step. Findings of EMNLP 2026, pages 48554–48571. DOI: 10.18653/2026.findings-emnlp.2268 · BibTeX