Trajectory matching trains discrete flows on structured action sequences
Original title: Trajectory Matching: Discrete Flow Matching with Structured Actions
Authors report MOSES validity rising from 92.8% to 99.2% versus factorized discrete flow matching, with transfer without retraining.
Trajectory matching conditions discrete flow matching on a full action sequence between endpoints rather than independent component-wise edits. The paper says the objective recovers the marginal flow and strictly subsumes factorized discrete flow matching, including size-changing and prerequisite actions. On tree and molecular graph generation, the authors report matching or beating factorized baselines and transfer to substructure extension and linking.
- Endpoint conditioning makes rates hard for non-factorizable actions.
- Actions may change state size or depend on earlier steps.
- Authors say the method recovers the marginal flow.
- An interactive explanation is posted on the author's blog.
From the original article
1 min read at arxiv.orgThe natural units of structured discrete generation are often not independent single-component edits but richer actions: attaching a fragment to a molecule, inserting a node together with an incident edge, or growing a hierarchy by adding a child. Yet most existing discrete flow matching methods (DFM) are restricted to independent component-wise edits over fixed-size states.
By Niklas Kemper · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.