Reproduce “Trajectory Matching: Discrete Flow Matching with Structured Actions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Niklas Kemper, Marten Lienen, David Lüdke, Stephan Günnemann
Trajectory matching extends discrete flow matching by conditioning training on valid action sequences between source and target, then learning when to take each action. This supports multi-component operations and prerequisites while retaining the flow-matching framework. On trees, child-attachment actions keep all reachable states valid, with 100% validity at 10 Euler steps versus about 8% for DeFoG; this result is task-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.