Berkeley scales online OT minibatch couplings to stabilize flow matching training
UCBerkeley · hf · 2026-10-15
UC Berkeley's Peter Chen and Sergey Levine tackle stability in scaling flow models. Independent source-data pairing introduces heavy stochasticity, and minibatch optimal transport (OT) couplings help but are limited by small batches, while scaling couplings per-iteration is computationally prohibitive. They develop a scalable online entropic OT approximation enabling much larger couplings, and close the theoretical gap by proving strict reductions in full stochastic-gradient variance and integration error versus independent pairing, with gains growing as the online minibatch grows. Experiments on single-cell gene interpolation-flow and image-flow confirm improved stability and final generation quality.
More from Research
- Tencent ARC's Player-1 learns general game control from synthetic experience — _akhaliq · 2026-10-15
- Claim: $2,000 of AI compute cracked the unsolved Longest Path problem, unverified — Regular_Post9884 · 2026-10-15
- Kornia releases kornia-slam: open-source Rust SLAM with mono/stereo/IMU support — edgarriba · 2026-10-15
- Tree packing lands in TRL AsyncGRPO, cutting agentic RL trainer compute 26x — SergioPaniego · 2026-10-15
- ProbabilityBench: exact ground-truth uncertainty evaluation across 16 LLMs — ishidalab · 2026-10-15
- Vela: spline-based continuous trajectories scale vision-language-action models — Yifan Li · 2026-10-15