CrossVLA: Cross-Paradigm VLA Optimization

• CrossVLA is a cross-paradigm framework that unifies discrete-token and continuous-action policies for Vision-Language-Action models.
• The framework introduces a surrogate flow-matching log-probability estimator and compares LoRA versus DoRA, achieving notable improvements in success rates.
• Inference analysis reveals that the denoise loop dominates latency over prefix caching, guiding future efforts towards core process optimization.

Searching arXiv for papers on CrossVLA and closely related uses of the term. CrossVLA denotes a cross-paradigm framework for post-training and inference optimization in Vision-Language-Action (VLA) models, introduced as an empirical study spanning discrete-token autoregressive policies such as OpenVLA and continuous-action flow-matching policies such as $\pi_{0.5}$ (Liu, 21 May 2026). In that work, CrossVLA is defined by three contributions: a surrogate flow-matching log-probability estimator that allows Direct Preference Optimization (DPO) to operate on continuous-action backbones without probability-flow ODE integration; a parameter-efficient fine-tuning comparison between LoRA and DoRA under VLA DPO; and an inference-time analysis showing that the denoise loop dominates latency in flow-matching VLAs and sharply limits the utility of prefix K/V caching (Liu, 21 May 2026). The term has also appeared more broadly in adjacent literature as a shorthand for cross-modal or cross-view visual-language alignment, but the named method “CrossVLA” refers specifically to the VLA post-training framework of 2026 (Liu, 21 May 2026).

1. Definition and Scope

CrossVLA studies how preference alignment and inference optimization transfer across two major VLA architectural lineages: autoregressive token policies and flow-matching continuous-action policies (Liu, 21 May 2026). A VLA model is described as mapping $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$, and the framework is organized around a unified experimental and software interface that hides whether the underlying policy is token-autoregressive or flow-based (Liu, 21 May 2026).

The autoregressive branch is instantiated with OpenVLA, where each scalar action dimension is discretized into 256 bins and a 7-DoF action is represented as $7 \times 256$-bin tokens; the model uses a Llama-2 7B LLM with a fused DINO-SigLIP vision tower, and token log-probabilities are directly available (Liu, 21 May 2026). The continuous branch is instantiated with $\pi_{0.5}$, which emits continuous 7-DoF action chunks via flow matching, uses a PaliGemma vision-language encoder plus a dedicated 10-step ODE “action expert,” and does not natively provide tractable chunk log-probabilities (Liu, 21 May 2026).

Within the broader literature, “CrossVLA” has also been used descriptively rather than as a method name. LexVLA frames visual-language alignment as a shared lexical space and explicitly characterizes its alignment space from a “CrossVLA” lens (Li et al., 2024). “Visual-Text Cross Alignment” presents a local region-to-description scoring method for zero-shot vision-LLMs and describes this as visual-language cross alignment (Li et al., 2024). These works are conceptually related, but they address retrieval or zero-shot classification rather than VLA post-training. This suggests that the term has developed both a narrow sense—CrossVLA as a specific VLA framework—and a broader sense referring to structured cross-modal alignment.

2. Cross-Paradigm Formulation

CrossVLA is organized around a minimal protocol with methods policy_logp, policy_logp_with_ref, policy_sample, encode_obs, and sample_actions, allowing the same DPO and benchmarking code to be applied to OpenVLA and $\pi_{0.5}$ (Liu, 21 May 2026). For token-autoregressive models, policy_logp is the sum of teacher-forced token log-probabilities. For flow-matching models, policy_logp is provided by a surrogate because the true log-density would require probability-flow ODE integration and Jacobian computation (Liu, 21 May 2026).

The training pipeline begins from public supervised fine-tuning checkpoints: OpenVLA-7B LIBERO-finetuned per-suite models and a $\pi_{0.5}$ LIBERO SFT checkpoint converted to PyTorch (Liu, 21 May 2026). These checkpoints serve as frozen reference policies in DPO. Preference data are then collected as approximately 200 chosen/rejected chunk pairs per LIBERO suite by rolling out the SFT policy with injected action noise, with the noise standard deviation ramped from 0.1 to 0.4 (Liu, 21 May 2026). Post-training uses DPO with either LoRA or DoRA adapters, and evaluation is conducted on the LIBERO 4-suite benchmark: Spatial, Object, Goal, and Long-horizon, with 10 tasks per suite, 5 trials per task, and 3 seeds for the main DoRA analysis (Liu, 21 May 2026).

The framework also contributes a multi-view + temporal projection head trained on 6000 LIBERO frames, but this component is orthogonal to DPO and inference optimization (Liu, 21 May 2026). A plausible implication is that CrossVLA is intended as an implementation-level laboratory for VLA methods rather than as a single monolithic policy architecture.

3. Surrogate Log-Probability for Flow-Matching DPO

The central technical obstacle addressed by CrossVLA is that DPO requires $\log p_\theta(c \mid \text{obs})$ and $\log p_{\text{ref}}(c \mid \text{obs})$ for preferred and rejected action chunks, whereas flow-matching policies define a conditional velocity field rather than an explicit closed-form density (Liu, 21 May 2026). In $\pi_{0.5}$, actions are generated by integrating a 10-step denoise loop from noise $x_0$ to action $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$0, with the model predicting $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$1 along the interpolation path $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$2 (Liu, 21 May 2026).

CrossVLA introduces a surrogate chunk log-probability based on flow-matching velocity-prediction error: $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$3 where $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$4, $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$5, and the evaluation times are $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$6 plus an additional stochastic perturbation per step (Liu, 21 May 2026). The justification is empirical and analogical rather than exact: lower MSE is treated as a higher log-probability proxy, with the missing proportionality factors absorbed into the DPO temperature $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$7 (Liu, 21 May 2026).

The DPO loss is then written using exact log-probabilities for OpenVLA and surrogate log-probabilities for $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$8: $$(\text{image(s)}, \text{language instruction}, \text{proprioceptive state}) \rightarrow \text{robot action}$$9 with

$7 \times 256$0

and $7 \times 256$1 (Liu, 21 May 2026). Training uses AdamW with learning rate $7 \times 256$2, batch size 1, 500 steps, and warmup 100 (Liu, 21 May 2026).

Empirically, the surrogate is reported to produce stable DPO training curves for $7 \times 256$3, with monotonic loss decrease and positive chosen–rejected margin growth, while preserving saturated SFT performance on Spatial and Object (Liu, 21 May 2026). The reported suite-level results are unchanged at 100% for Spatial and 98% for Object after LoRA+DPO, indicating non-degenerate optimization rather than measurable headroom gains under those settings (Liu, 21 May 2026).

4. Parameter-Efficient Post-Training: LoRA versus DoRA

CrossVLA compares Low-Rank Adaptation (LoRA) and Weight-Decomposed Low-Rank Adaptation (DoRA) as the parameter-efficient substrate for VLA DPO (Liu, 21 May 2026). LoRA adds a rank-$7 \times 256$4 update $7 \times 256$5 to a frozen pretrained matrix $7 \times 256$6, while DoRA further decomposes the effective weight into magnitude and direction: $7 \times 256$7 where $7 \times 256$8 is a learnable per-output-channel magnitude vector (Liu, 21 May 2026).

On OpenVLA-7B with rank $7 \times 256$9, LoRA-r32 has 33.55M trainable parameters and 17.93 GB peak GPU memory at evaluation, while DoRA-r32 has 34.08M trainable parameters and 26.17 GB peak GPU memory (Liu, 21 May 2026). The difference in trainable parameter count is marginal, but DoRA incurs higher runtime memory because the effective weight is materialized during forward passes (Liu, 21 May 2026).

The main quantitative comparison is reported over the LIBERO 4-suite. Relative to the reproduced SFT baseline, DoRA improves mean success from 62.75% to 73.2%, a mean gain of +10.4 percentage points across 600 trials and 3 seeds (Liu, 21 May 2026). Per-suite gains are +20.0 on Object, +11.0 on Long-horizon, +8.0 on Goal, and +2.7 on Spatial (Liu, 21 May 2026). The Object suite is especially notable: SFT achieves 56%, DoRA reaches 76.0% pooled over 150 trials, and each of the three seeds records exactly 38/50 successes, yielding zero seed variance on that suite (Liu, 21 May 2026).

The reported table is central to the framework’s practical conclusions:

Suite	SFT (ours)	DoRA 3-seed pool	$\pi_{0.5}$0 vs SFT
Spatial	72%	74.7% (112/150)	+2.7 pp
Object	56%	76.0% (114/150)	+20.0 pp
Goal	70%	78.0% (117/150)	+8.0 pp
Long-horizon	53%	64.0% (96/150)	+11.0 pp
Mean	62.75%	73.2%	+10.4 pp

The paper attributes these gains to the narrow-domain nature of LIBERO post-training: visual grounding is already strong in the SFT checkpoint, and adaptation often requires controlled magnitude adjustments rather than large directional changes in feature space (Liu, 21 May 2026). This suggests that DoRA is especially effective when the task distribution is close to the pretrained policy’s support and reproducibility is a priority.

5. Inference-Time Anatomy and Cache Limits

CrossVLA’s inference analysis targets the flow-matching branch, specifically $\pi_{0.5}$1, whose sample_actions routine first computes a PaliGemma prefix and then runs a 10-step denoise loop through the action expert (Liu, 21 May 2026). On LIBERO Spatial using an H20 GPU, the measured latency per sample_actions call is approximately 5 ms for image preprocessing and tokenization, 60 ms for embed_prefix plus the PaliGemma prefix forward pass, and 220 ms for the denoise loop, for a total of roughly 280 ms (Liu, 21 May 2026). The denoise loop therefore accounts for 78.6% of latency, while the prefix accounts for 21.4% (Liu, 21 May 2026).

This cost decomposition yields a hard upper bound: prefix-only acceleration cannot exceed the prefix share of runtime, so prefix K/V caching cannot deliver more than about a 21% speedup even under ideal reuse (Liu, 21 May 2026). CrossVLA evaluates two cache strategies. Chunk-level caching reuses entire action chunks when the visual embedding cosine similarity exceeds 0.95; on LIBERO Spatial, this produces an 82.1% chunk reuse rate and mean similarity 0.99, yet wall time increases from 1258 s to 1796 s over 50 trials and success drops from 50/50 to 40/50 (Liu, 21 May 2026). The overhead is attributed to similarity computation, CPU pool management, and tensor copies, while failure is attributed to stale action reuse under slight environment divergence (Liu, 21 May 2026).

Token-level prefix K/V caching, implemented by monkey-patching sample_actions to reuse the prefix K/V while recomputing the action expert, performs worse. Under realistic thresholds, success collapses to 0–80%, and the only 100% success case is a no-hit sanity check with a 0.999 similarity threshold (Liu, 21 May 2026). The study concludes that prefix caching, although effective for autoregressive VLAs, is structurally mismatched to flow-matching VLAs because the dominant compute resides in the denoise loop rather than the prefix (Liu, 21 May 2026).

The explicit recommendation is therefore to accelerate the denoise loop itself, for example by distilling 10-step flows into 1–4 steps via consistency or progressive distillation; concurrent SnapFlow is cited as validating this direction (Liu, 21 May 2026). This is one of the clearest negative results in the CrossVLA study: caching is not merely less helpful for flow VLAs, but can be counterproductive under benchmark conditions.

6. Auxiliary Representation Learning and Benchmarking Context

Beyond DPO and inference profiling, CrossVLA trains a small multi-view + temporal projection head on top of OpenVLA’s frozen SigLIP vision encoder (Liu, 21 May 2026). The head maps a 1152-dimensional pooled SigLIP feature through Linear(1152, 512) -> [GELU](https://www.emergentmind.com/topics/gaussian-error-linear-unit-gelu) -> Linear(512, 128) -> L2 normalize, with only 656K trainable parameters (Liu, 21 May 2026). Training uses 6000 frames drawn from LIBERO RLDS, comprising 4 suites $\pi_{0.5}$2 50 episodes $\pi_{0.5}$3 30 anchor times, with dual-stream InfoNCE over synchronized agent-view and wrist-view images and temporal consistency between agent views separated by $\pi_{0.5}$4 (Liu, 21 May 2026).

Optimization uses AdamW with learning rate $\pi_{0.5}$5, cosine decay, batch size 128, equal weights $\pi_{0.5}$6, and temperature $\pi_{0.5}$7, for 10 epochs in roughly 30 minutes on a single H20-3e (Liu, 21 May 2026). The resulting representation achieves 99.5% Recall@1 for same-task nearest-neighbor retrieval, 99.9% Recall@5, and 99.95% Recall@10, compared with a 2.75% random baseline at Recall@1 (Liu, 21 May 2026). Same-episode Recall@1 is 91.4%, and same-task with $\pi_{0.5}$8 reaches 92.4% Recall@1 (Liu, 21 May 2026).

This component is positioned as a downstream initialization rather than a control policy improvement. The paper releases the projection head checkpoint proj_head.pt and suggests uses such as same-task retrieval, task indexing, or auxiliary initialization for SigLIP-based VLAs (Liu, 21 May 2026). No direct improvement in robot success rate from this head is reported.

The benchmarking context is entirely centered on LIBERO simulation. For seed-42 SFT reproduction, OpenVLA records 72% on Spatial, 56% on Object, 70% on Goal, and 53% on Long-horizon, while $\pi_{0.5}$9 reaches 100.0%, 98.0%, 100.0%, and 94.0% respectively, closely matching its original paper (Liu, 21 May 2026). This asymmetry is important for interpreting CrossVLA results: the AR branch provides room for DPO gains, whereas the flow branch is largely saturated in the reported suites.

7. Position in the Broader “CrossVLA” Landscape

CrossVLA is situated within a heterogeneous body of work using similar terminology for different forms of cross-modal alignment. LexVLA replaces dense latent alignment with a unified lexical representation over a shared vocabulary and explicitly presents itself as a “Lexical CrossVLA” paradigm in which images and texts are aligned as sparse token-weight vectors (Li et al., 2024). Its alignment space is interpretable, uses DINOv2 and LLaMA 2 with modality-specific codebooks, and introduces an overuse penalty to suppress meaningless token activation (Li et al., 2024). In that setting, “CrossVLA” refers to structured vision-language alignment rather than robot control.

“Visual-Text Cross Alignment” likewise studies visual-language alignment, but in a training-free zero-shot classification setting where local image crops are matched against multiple LLM-generated class descriptions using a weighted similarity matrix (Li et al., 2024). Its score function,

$\pi_{0.5}$0

is designed to replace global image-text similarity with local cross alignment between visual regions and fine-grained text (Li et al., 2024). Again, the topic is alignment rather than action.

The term also appears in embodied learning beyond the 2026 CrossVLA paper. X-VLA is a cross-embodiment VLA that uses embodiment-specific soft prompts within a shared Transformer and flow-matching policy, with 0.04% unshared parameters per data source and a 0.9B-parameter backbone trained on approximately 290K episodes across 7 platforms (Zheng et al., 11 Oct 2025). VLA-Pro frames cross-task generalization as retrieval and fusion of task-specific LoRA adapters as procedural memories, reporting up to a 207% relative improvement in simulation and an increase in real-world success rate from 5.8% to 65.0% (Si et al., 28 May 2026). CrossVL addresses cross-view vision-language detection through Complexity-Aware Pathway Aggregation and Paired Curriculum Learning, improving Florence-2’s aerial mAP from 58.66% to 61.03% on MAVREC (Liu et al., 10 May 2026).

These works indicate that “CrossVLA” now spans at least four distinct research directions: cross-paradigm post-training for robot policies, cross-modal lexical alignment, cross-view detection, and cross-embodiment or cross-task transfer in robotics. The named method CrossVLA remains the post-training and inference framework of 2026 (Liu, 21 May 2026), but the surrounding literature suggests a broader organizing theme: explicit mechanisms for transferring alignment, adaptation, or optimization across mismatched paradigms, views, embodiments, or task families.

8. Limitations and Prospective Directions

The CrossVLA study is explicit about several limitations. Only two backbones are fully tested, since a planned third backbone based on Qwen3-VL flow matching or Spirit v1.5 was dropped due to runtime issues (Liu, 21 May 2026). GRPO is evaluated only on a single Spatial suite, where it improves SFT by +2 percentage points but underperforms DPO at the available compute budget (Liu, 21 May 2026). Flow-matching DPO experiments are incomplete on Goal and Long-horizon because of GPU pod instability, so cross-paradigm validation on $\pi_{0.5}$1 is limited to Spatial and Object (Liu, 21 May 2026). The multi-view projection head is evaluated only by retrieval metrics, not by downstream control improvement (Liu, 21 May 2026). All experiments are conducted in LIBERO simulation, with no real-robot validation (Liu, 21 May 2026).

The OpenVLA SFT reproduction also uses default decoding temperature 1.0 rather than the settings used in the original OpenVLA paper, leading to lower absolute baseline numbers on some suites (Liu, 21 May 2026). This matters because the magnitude of DoRA improvements is measured against those reproduced baselines, not necessarily against the strongest published decoding configuration.

The generality claims are correspondingly circumscribed. The surrogate log-probability is argued to transfer to other flow-matching or diffusion-like control models because it depends only on evaluating MSE-style denoising or velocity loss at a small set of timepoints (Liu, 21 May 2026). The DoRA-versus-LoRA conclusion is suggested to generalize to other narrow-domain post-training settings in robotics and beyond (Liu, 21 May 2026). The inference analysis is argued to apply to any VLA whose runtime is dominated by a denoising or ODE loop rather than by prefix encoding (Liu, 21 May 2026).

Taken together, CrossVLA establishes three specific conclusions. First, DPO can be extended beyond autoregressive VLA policies by replacing intractable flow log-densities with a practical surrogate based on flow-matching error (Liu, 21 May 2026). Second, DoRA is a strong default PEFT mechanism for VLA preference optimization in narrow domains, with a mean +10.4 percentage point improvement over reproduced OpenVLA SFT across the LIBERO 4-suite (Liu, 21 May 2026). Third, inference optimization for flow-matching VLAs must focus on the denoise loop rather than on prefix caching, because the denoise loop dominates 78.6% of latency and constrains prefix-only acceleration to a 21% ceiling (Liu, 21 May 2026).