Research line · emerging
Transition matching
Model generation as discrete-time Markov transitions over a continuous state, unifying flow matching and continuous-token autoregression.
What defines membership
Flow matching and autoregression are not rival paradigms but two settings of one transition-kernel formulation.
How the line developed
Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.
The idea
Transition Matching · 2025-06core
Recasts generation as discrete-time Markov transitions over a continuous state, unifying flow matching and continuous-token autoregression.
Evidence
Demystifying TM · 2025-10core
Independent proof that transition matching achieves strictly lower KL divergence than flow matching at finite step counts.
Improvement
TM Design Space · 2025-12strong-followup
Sweeps the head design across 56 models at 1.7B scale to find which configurations actually work.
TM Distillation · 2026-01emerging
Extends transition matching to distilling video generation. Its own ablation shows plain TM pretraining nearly matches the full TM-MeanFlow objective, and it claims only the narrow video result, not a general validation of the paradigm.
What it gets right
- Offers a single formalism where the field currently has two separate camps
- From the group that introduced flow matching, so the theoretical framing is credible
Where it is weak
- One paper deep; no independent replication yet
- Unclear whether the unification yields practical gains or mainly conceptual clarity