← News

ResearchSources: arXiv · 18 Nov 2026

ToR-OT: transformer learns to predict optimal transport plans as token sequences

An arXiv paper introduces ToR-OT, a transformer model that decodes sparse transport plans directly as sequences, avoiding test-time optimization on the discrete problems studied.

In plain words

Optimal transport is a maths problem about moving between two probability distributions at the lowest cost, and it is usually solved with classical optimization solvers. The authors built a transformer, a type of neural network, that instead generates the answer as a short sequence of tokens. Because exact solutions are sparse, the model only needs to output the nonzero moves.

What was announced

The paper, posted on arXiv, describes ToR-OT as a transformer-based model trained on a large synthetic corpus of optimal transport instances. The authors recast the task as structured sequence generation: rather than solving an optimization problem at inference time, the model autoregressively predicts a discrete transport plan, emitting only the nonzero entries that exact discrete solutions contain.

According to the abstract, a single pretrained model generalized zero-shot to unseen distribution pairs and heterogeneous cost metrics without retraining, and on low-dimensional discrete problems it produced accurate approximate plans that compared favorably with per-instance neural OT baselines. The authors stress that ToR-OT is not yet a replacement for classical solvers, but a step toward understanding how autoregressive transformers can learn the structure of constrained optimization problems.

What it does not show

  • The evidence is the paper's abstract only; the reported generalization and accuracy claims have not been independently verified, and results are limited to low-dimensional discrete problems as described.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.