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

2026-05-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

CrossVLA presents an empirical study on post-training and inference optimization for Vision-Language-Action (VLA) models, addressing both discrete-token autoregressive and continuous-action flow-matching architectures. Key contributions include a surrogate flow-matching log-probability estimator enabling Direct Preference Optimization (DPO) on continuous-action backbones without prohibitive ODE integration. The study also demonstrates that DoRA, a parameter-efficient fine-tuning method, improves OpenVLA SFT by a mean +10.4 pp across the LIBERO 4-suite, with specific gains like +20.0 pp on Object tasks and zero seed variance. Furthermore, an inference-time analysis reveals the denoise loop dominates 78.6% of `sample_actions` latency in flow-matching models, limiting prefix-K/V caching acceleration to 21% and showing caching degrades success rates. A multi-view + temporal projection head pretrained on 6000 LIBERO frames is also released, achieving 99.5% k-NN recall@1.

Key takeaway

For Machine Learning Engineers optimizing Vision-Language-Action models, you should prioritize DoRA over LoRA for post-training with DPO, as it yields substantial performance improvements, such as +20.0 pp on Object tasks. When tackling inference latency in flow-matching VLAs, focus your efforts on optimizing the denoise loop, which accounts for 78.6% of `sample_actions` time, rather than K/V caching, which offers limited gains and can degrade success rates.

Key insights

CrossVLA enables DPO for continuous-action VLAs and optimizes inference, showing DoRA's superiority and denoise loop's latency dominance.

Principles

• DPO can be adapted for continuous-action VLAs using a tractable surrogate log-probability.
• DoRA generally outperforms LoRA for VLA DPO across various task families.
• Inference optimization for flow-matching VLAs should target the denoise loop, not prefix caching.

Method

CrossVLA introduces a negative-MSE flow-matching surrogate log-probability for DPO on continuous-action VLAs, and empirically compares PEFT layers (LoRA vs DoRA) and inference caching strategies.

In practice

• Apply DoRA for VLA DPO to achieve significant performance gains.
• Focus VLA inference optimization on the denoise loop.
• Utilize the released multi-view projection head for task retrieval.

Topics

• Vision-Language-Action Models
• Direct Preference Optimization
• Parameter-Efficient Fine-Tuning
• DoRA
• Inference Optimization
• Flow-Matching Models

Code references

• lz-googlefycy/vla-lab

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Fine-Tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation — LoRA/DoRA fine-tuning effectively adapts large world models like Cosmos Predict 2.5 for domain-specific robot video generation.
• You Only Index Once: Cross-Layer Sparse Attention with Shared Routing — Cross-layer sparse attention (CLSA) shares KV cache and routing indices across layers to improve LLM inference efficiency and quality.
• AVIS: Adaptive Test-Time Scaling for Vision-Language Models — AVIS adaptively scales visual context and reasoning for VLMs, significantly reducing inference cost while improving accuracy.
• ProxyKV: Cross-Model Proxy Pruning for Efficient Long-Context LLM Inference — ProxyKV uses an asynchronous small-model proxy to prune KV caches, achieving high accuracy and significant speedups for long-context LLM inference.
• Towards Selective Depth Mixing: Attention Residuals for Stable, Reusable Features in Deep Transformers — Novel architectural and inference techniques enhance LLM efficiency and performance by optimizing residual connections, KV cache management, and sampling.
• LARA: Latent Action Representation Alignment for Vision-Language-Action Models — LARA jointly optimizes Latent Action Models and Vision-Language-Action models via representation alignment for improved robot control.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.