CrossVLA: Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models

Abstract: Vision-Language-Action (VLA) models have rapidly converged on a small set of architectural patterns: discrete-token autoregression (e.g. OpenVLA) and continuous-action flow-matching (e.g. pi-0.5). Yet preference alignment via Direct Preference Optimisation (DPO) -- the de-facto post-training step in LLMs -- has been studied almost exclusively on autoregressive VLAs. We present CrossVLA, an empirical study of cross-paradigm VLA post-training. Three contributions: (i) a surrogate flow-matching log-probability estimator that lets DPO operate on continuous-action backbones without probability-flow ODE integration; (ii) a head-to-head comparison of LoRA and DoRA as the parameter-efficient layer for VLA DPO, finding DoRA improves over OpenVLA SFT by a mean +10.4 pp across LIBERO 4-suite (600 trials, 3 seeds) -- per-suite +20.0 Object, +11.0 Long-horizon, +8.0 Goal, +2.7 Spatial -- with zero seed variance on Object (38/50 on each of 3 seeds); (iii) an inference-time anatomy showing the denoise loop dominates 78.6% of sample_actions latency and prefix-K/V caching a la VLA-Cache caps at a 21% acceleration ceiling -- both chunk-level and token-level cache strategies degrade success rate to 0-80% in our benchmarks. We further pretrain a multi-view + temporal projection head on 6000 LIBERO frames, achieving 99.5% k-NN recall@1 for same-task retrieval (36x over random), available as a downstream initialisation. All code, ckpts, training logs, and reproduction scripts are open at https://github.com/lz-googlefycy/vla-lab.

• The paper presents a novel surrogate negative-MSE log-probability for DPO on flow-matching VLA models, enabling stable cross-paradigm post-training.
• It adapts DoRA for parameter-efficient fine-tuning, achieving up to +10.4pp gains over SFT baselines on the LIBERO benchmark.
• It reveals that traditional caching strategies are ineffective for flow-matching models, guiding future research towards denoise-loop-targeting accelerations.

Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models: An Expert Analysis

Introduction and Context

The paper "CrossVLA: Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models" (2605.21854) systematically investigates post-training optimization and inference strategies across heterogeneous Vision-Language-Action (VLA) model architectures. The work aims to bridge the gap between discrete-token autoregressive policies and continuous-action flow-matching backbones within the context of VLA models for embodied robotics, principally evaluated on the LIBERO benchmark.

VLAs have coalesced around two dominant architectural families: models like OpenVLA (autoregressive, discretized action tokens) and $\pi_{0.5}$ (continuous, flow-matching, ODE-based action generation). The paper addresses the underexplored area of cross-paradigm post-training, especially the extension of Direct Preference Optimization (DPO)—widely successful in LLMs—to these varied VLA frameworks.

Methodological Contributions

The primary technical contributions are:

1. Surrogate Log-Probability for Flow-Matching Backbones: The authors introduce a negative-MSE-based surrogate for the log-probability required by DPO on continuous-action VLA models. Traditional DPO requires tractable $\log p_\theta(\text{chunk}|\text{obs})$, which is intractable for flow-matching policies due to the need for probability-flow ODE integration and associated Jacobians. The surrogate allows efficient and stable DPO training on models like $\pi_{0.5}$ without undermining the theoretical soundness of preference optimization.
2. Parameter-Efficient Fine-Tuning (PEFT) Adapter Selection: DoRA (Weight-Decomposed Low-Rank Adaptation), originally from LLMs, is adapted as a PEFT method for VLA DPO. DoRA decouples magnitude and direction in weight updates, in contrast to LoRA's joint update. Empirically, DoRA yields a striking mean +10.4 percentage point (pp) gain over OpenVLA SFT across LIBERO’s four suites (Object: +20.0pp, Long-horizon: +11.0pp, Goal: +8.0pp, Spatial: +2.7pp), with zero seed variance on the Object suite, indicating reproducibility and robust optimization in narrow-distribution learning scenarios.
3. Inference Latency Anatomy and KV-Cache Analysis: The study provides a detailed breakdown of inference-time costs in flow-matching VLAs, showing that the denoise loop accounts for 78.6% of sample_actions latency, rendering classic KV/prefix caching strategies (such as VLA-Cache) from the autoregressive setting ineffective—the acceleration ceiling is capped at 21%. Both chunk-level and token-level cache strategies are shown to degrade task success rates substantially while offering little speedup.
4. Multi-View + Temporal Contrastive Pretraining of Visual Features: A projection head atop a frozen SigLIP encoder is pretrained using both multi-view and temporal contrastive InfoNCE objectives. This produces a highly discriminative 128-dimensional feature, achieving 99.5% k-NN recall@1 for same-task retrieval on 6000 LIBERO frames, a $36\times$ improvement over random retrieval.

All code and artifacts are made public, ensuring full reproducibility.

Empirical Findings

Surrogate Log-Likelihood for DPO on Flow-Matching Architectures

The negative-MSE surrogate is validated as an effective drop-in replacement for intractable log-likelihood computations in flow-matching policies. DPO with this surrogate retains stable training dynamics, and produces non-degenerate preference-aligned policies. On LIBERO, SFT baselines are matched or closely reproduced; in SFT-saturated regimes (Spatial, Object), DPO offers no further gains, consistent with ceiling effects.

DoRA vs LoRA: Effectiveness and Stability

DoRA consistently outperforms LoRA in both average gain and variance reduction, especially notable on the Object suite (zero variance across three seeds). DoRA’s decoupling of magnitude and direction proves especially valuable in narrow-distribution finetuning scenarios typical of SFT-to-preference-alignment transitions in robotics. Furthermore, DoRA demonstrates tighter per-seed success distributions, mitigating variance and sensitivity to "lucky" seeds that plague LoRA.

Inference Optimization: Caching Strategies

Empirical latency studies expose the inadequacy of prefix K/V caching in the flow-matching context—since most computation cost is incurred in the denoise loop, not the vision-language prefix forward pass. Neither chunk-level nor token-level cache strategies realize meaningful acceleration and both impair control performance. The negative result is structural, pointing to consistency model distillation and denoise-loop-targeting accelerations as the sole productive avenues for fast inference in flow-matching VLAs—a direction concurrently validated by SnapFlow (Luan et al., 7 Apr 2026).

Multi-View + Temporal Pretraining

The contrastive projection head converges rapidly, closing 92.5% of the loss gap relative to the random baseline in under 500 steps. The k-NN recall@1 (same-task = 99.5%) evidences the practical utility of this representation for downstream robotic task retrieval. Usage as initialization for end-to-end policy finetuning remains unexplored and is highlighted as an immediate direction for follow-up.

Theoretical and Practical Implications

Paradigm-Independent Post-Training

By validating DPO with appropriate log-probability estimators across discrete and continuous action-generators, the work demonstrates the extension of preference alignment protocols from LLMs to the entire VLA architecture landscape. This ability to use one post-training method with minimal architectural plumbing is practically valuable for unified system design and benchmarking.

Parameter-Efficient Adapter Suitability

The findings provoke further investigation into the relationship between adapter architecture (e.g., magnitude-direction decoupling) and task distributional properties. The suggestion to diagnose suite-specific overfitting risk as a heuristic for adapter selection is a practical guideline for roboticists deploying VLAs in the wild.

Inference Optimization: Negative Results as Guidance

The structural argument dissecting why prefix caching is unfit for flow-matching guides future research directions toward denoise-loop compressions (consistency model distillation), already making headway via concurrent works.

Representation Learning for Downstream Bootstrapping

The extremely high retrieval performance of the contrastively pretrained projection head invites questions on the impact of such vectored representations as initializations or priors for SFT or DPO stages in VLA pipelines—potentially upgrading data efficiency and transfer capabilities.

Limitations and Future Directions

The work is transparent regarding remaining limitations: incomplete multiseed LoRA coverage, lack of evaluation on real robot hardware, lack of downstream policy evaluation for pretrained representations, and incomplete cross-paradigm experiments on the hardest suites. Extending DoRA+DPO to flow-matching backbones is an immediate follow-up. Real-world evaluation, comprehensive cross-algorithm grids, and investigating downstream impacts of strong visual pretraining are all logical future steps.

Conclusion

CrossVLA demonstrates practical post-training and inference procedures that transfer across major VLA architectural paradigms, defines structural boundaries for inference speed-up via caching in flow-matching policies, and empirically establishes the advantage of DoRA for narrow-distribution, parameter-efficient policy adaptation. The work advances the toolkit for unified preference-aligned VLA policy development and offers empirically validated prescriptions for both representation learning and inference acceleration (2605.21854).