CrossVLA Improves VLA Post-Training with Cross-Paradigm DPO

Researchers introduced CrossVLA, a study on cross-paradigm post-training for Vision-Language-Action (VLA) models. They developed a surrogate flow-matching log-probability estimator enabling Direct Preference Optimisation (DPO) on continuous-action backbones without probability-flow ODE integration. Comparing LoRA and DoRA for parameter-efficient VLA DPO, DoRA improved over OpenVLA SFT by a mean +10.4 percentage points across LIBERO 4-suite (600 trials, 3 seeds), with per-suite gains of +20.0 Object, +11.0 Long-horizon, +8.0 Goal, and +2.7 Spatial, and zero seed variance on Object (38/50 on each). The work addresses the gap where DPO had been studied almost exclusively on autoregressive VLAs like OpenVLA, extending it to continuous-action flow-matching models like pi-0.5.