Diffusion-Trained Backbones for Flow-Matching Vision-Language-Action Policies

1Abstract

Vision-Language-Action (VLA) policies pair a pretrained multimodal backbone with a continuous action generator such as a flow-matching action expert. Most backbones inherit an autoregressive (AR) next-token pretraining prior, yet robot action chunks are block-structured rather than left-to-right. We ask whether a diffusion-trained bidirectional backbone produces better action-token hidden states than a scale-matched AR backbone when both drive the same flow-matching action expert. We isolate the pretraining prior with a matched pair, Dream-7B (discrete diffusion, initialized from Qwen2.5-7B) versus Qwen2.5-7B (AR), holding architecture, initialization, action expert, dataset, and control stack fixed, and evaluate entirely in simulation on the LIBERO benchmark with both policy-level success and representation-level probing. On the matched C−D contrast (single seed, 500 rollouts/suite), the diffusion-trained backbone improves mean LIBERO success by +5.9 points (82.2 vs. 76.3). The gain is concentrated exactly where the autoregressive backbone is weakest: large and statistically significant (disjoint 95% Wilson intervals, p<0.001) on the long-horizon (+12.0) and object (+11.2) suites, and a tie on the near-ceiling spatial and goal suites. Linear probing of the two backbones' action-token features, however, finds them equally decodable for low-level action, state, and gripper targets (|C−D|≤0.01): the behavioral advantage is not reducible to the linear decodability of immediate control signals, which rules out the simplest explanation and leaves the underlying mechanism to future probing. We report the study as a controlled, single-seed result and discuss its limitations, including seed variance and the objective-vs-corpus confound inherent to off-the-shelf backbones.

Index terms: vision-language-action policies, flow matching, diffusion models, imitation learning, representation probing.

2Introduction

State-of-the-art VLA policies pair a pretrained vision-language model (VLM) backbone with a flow-matching action expert that turns the backbone's hidden states into continuous action chunks, as in π0 and π0.5 [2,3,4]. The action expert generates a whole block of future actions at once, refining them iteratively and attending across the block bidirectionally, a generative process much closer to denoising than to left-to-right prediction. The backbone that feeds it, however, is almost always an autoregressive VLM, pretrained to predict the next token causally. Does the backbone's pretraining objective, the prior it brings to the action-token hidden states, matter for the policy, and would a backbone pretrained by denoising suit a flow-matching expert better than a matched autoregressive one?

The question has been hard to ask cleanly for two reasons. First, until recently there was no diffusion-pretrained language model at the scale of the autoregressive backbones used in VLAs; discrete-diffusion LMs such as Dream-7B [5], initialized from Qwen2.5-7B [6], now provide an architecturally identical counterpart whose defining difference is a discrete-diffusion pretraining objective. Second, the inference-time attention mask is easy to confuse with the pretraining prior: the π0.5 backbone is in fact a prefix-LM whose action block is already attended bidirectionally, so the causal-vs-bidirectional distinction at inference is not the variable of interest. These observations make a controlled comparison possible: hold the action expert, vision encoder, data, and adaptation recipe fixed, and swap only the backbone between a diffusion-trained model (Dream-7B) and its autoregressive initializer (Qwen2.5-7B).

We run exactly this comparison on LIBERO. The diffusion-trained backbone yields the better policy (+5.9 points of mean success and +12.0 on the long-horizon suite, precisely where the autoregressive backbone is weakest) at matched scale, initialization, and everything downstream of the backbone, so the gap is attributable to the pretraining prior. Probing the two backbones' action-token features, however, finds them equally linearly decodable for low-level control targets: the behavioral advantage is real but does not reduce to the linear readability of immediate actions.

Our contributions are: (i) a scale/architecture-matched isolation of the diffusion pretraining prior for VLA backbones; (ii) a precise characterization of the π0.5 backbone as a prefix-LM (bidirectional action block), separating attention-mask topology from the pretraining prior; (iii) representation-level probing of action-token hidden states; and (iv) a controlled LIBERO evaluation with rollout-level statistics (Wilson intervals, two-proportion tests) and an explicit account of the study's confounds and single-seed scope.

3Related Work

Diffusion and flow policies. Generating actions by denoising or flow matching is now standard in imitation learning: Diffusion Policy [1] and 3D Diffusion Policy [13] show that iterative, block-structured action generation captures multimodal behavior better than regressing a single action. These methods build the denoiser on task-specific or from-scratch encoders; we inherit the flow-matching action head but ask what its backbone should be.

VLM backbones with flow-matching experts. π0 and π0.5 [2,3,4] scale this recipe by conditioning a flow-matching expert on a large pretrained VLM. In these systems the backbone is autoregressively pretrained and (in π0.5) used as a prefix-LM, so the action block is already bidirectional at inference; the pretraining objective of the backbone is not itself treated as a design choice. Our study takes that objective as the variable.

Diffusion-trained and multimodal-diffusion LMs. Discrete-diffusion LMs (Dream-7B [5], and encoder-decoder diffusion LMs such as DiffusionGemma [11]) and multimodal diffusion models like LaViDa [7] establish that denoising-pretrained backbones exist at VLM scale. Crucially, Dream is initialized from an autoregressive model (Qwen2.5-7B), giving a matched pair whose principal difference is the pretraining objective (with the corpus/compute caveat we bound in §VI).

Unified-backbone diffusion VLAs (closest work). A recent line folds action decoding into a diffusion backbone (Dream-VLA [8], LLaDA-VLA [9], Discrete Diffusion VLA [10]), reporting strong LIBERO results. These works differ from ours in three ways that matter for what can be concluded: they (a) do not hold a matched autoregressive backbone fixed as a control, so the pretraining-prior effect is confounded with architecture and training changes; (b) let the backbone generate actions directly rather than through a separate, fixed flow-matching expert, so representation quality and decoding are entangled; and (c) do not probe the learned representations. We instead keep one external flow-matching expert constant, swap only the backbone against its own AR initializer, and add representation-level probing, trading peak performance for a controlled, interpretable contrast.

Representation quality and action-token structure. Crossway Diffusion [12] shows auxiliary representation learning improves diffusion policies, and work on ordered/structured action tokens [14] highlights that how actions are laid out as tokens affects learning, both motivating our focus on the action-token hidden states the expert consumes and our attempt to measure their quality directly via probing.

4Method

4.1Backbone-swap with a shared flow-matching expert

Image/language/proprio/flow-time/noisy-action tokens are processed by the backbone, which contextualizes the action tokens; the hidden states at the action-token positions, H_action, are read by a flow-matching action expert that outputs the vector field over the action chunk. The flow objective is

Standalone vs. fused expert. π0.5's action expert is not a detachable head: it is a second transformer stream fused into the backbone's self-attention (per layer, backbone and expert queries/keys/values are concatenated and attended jointly, sharing RoPE and KV cache), which forces the expert to match the backbone's layer count, head geometry, and normalization. This fusion cannot be preserved when the backbone changes from Gemma to Qwen2.5/Dream. We therefore use a standalone flow expert, a small DiT-style transformer over the action tokens with flow-time injected via adaLN-Zero conditioning, that reads H_action and is decoupled from the backbone's attention. It is shared identically across variants A′/C/D; the only variant-specific parameter is a linear input adapter mapping backbone width to the common expert width (~2–4M of ~117M parameters), so the C−D contrast stays clean. Model A (stock π0.5, fused expert) is retained as a reference that reproduces published numbers, and A′ (Gemma backbone + the standalone expert) is the architecture-matched baseline.

4.2Model variants

The action expert, action space, flow objective, dataset, and control stack are identical across all rows; only the backbone/attention topology changes.

4.3Reading hidden states from a diffusion backbone

A diffusion-trained backbone could in principle be run for several internal refinement steps (k>1) before its hidden states are read, unrolling the denoising process the AR backbone lacks. To keep the C−D contrast about the pretraining prior and not about extra inference compute, we read both backbones with a single forward pass (k=1): C and D then do exactly the same amount of work per action, and any gap is due to the learned representation, not additional refinement. Multi-step refinement reads (k∈{2,4}) are a natural ablation left to future work.

5Experimental Setup

Evaluation is entirely in simulation on LIBERO [15] (single-arm Franka, 7-DoF), across four suites (Spatial, Object, Goal, Long / libero_10), 10 tasks per suite × 50 trials = 500 rollouts/suite. Backbones are ~7B (Dream-7B [5] / Qwen2.5-7B [6]); Model A is the stock π0.5 (Gemma) backbone. Vision is held identical across the standalone-expert variants (A′/C/D): the SigLIP tower from PaliGemma plus a fresh linear connector, so only the language backbone changes between C and D. Training forks the openpi pi05_libero recipe (30k steps, batch 32, LR 5×10⁻⁵, action horizon H=10, action dim padded to 32).

Adaptation differs by variant and is stated with each result: A and A′ are fully fine-tuned from pi05_base; C and D use an identical strengthened-LoRA recipe (rank 64, α=128 on the frozen 7B backbone, with the SigLIP tower, connector, input adapters, and standalone expert trainable), chosen because full fine-tuning of a 7B backbone with AdamW does not fit one 80GB GPU under DDP, and because freezing the base best preserves the pretraining prior the C−D contrast is designed to isolate. The diffusion backbone is read with a single forward pass (k=1), matching D exactly. Each training run used 2×A100 (~5 s/step). We report per-suite success rate with 95% Wilson confidence intervals (500 rollouts/suite) and a two-proportion test on the C−D contrast, with LIBERO-Long as the primary discriminator, plus representation probes (§5.3). A and A′ are averaged over 3 seeds; C and D are single-seed (n=1); the identifying C−D contrast is evaluated on the matched seed.

Reproducibility. The implementation extends OpenPI [4] (Apache-2.0); the backbones (Dream-7B, Qwen2.5-7B) and the LIBERO benchmark are public, and no private data is used. The standalone flow expert, per-variant training configs, and the probing pipeline are implemented as an openpi.diffusion_backbone package; code and configurations will be released.

6Results

6.1Stage 1: Model A baseline reproduction

Our Model A reproduces the published π0.5 LIBERO results to within ~1% on every suite (average 96.6 vs. 96.85), validating the training and evaluation harness before any backbone swap. This holds despite training at batch size 32 (vs. the published 256). On the long-horizon suite Model A slightly exceeds the published number (93.1 vs. 92.4).

6.2Main comparison: diffusion vs. matched AR backbone

The diffusion prior helps, and most where AR is weakest. On the identifying C−D contrast (matched seed, identical standalone expert, LoRA recipe, and vision), the diffusion-trained backbone improves mean success by +5.9 points. The gain is not uniform, and the pattern is the informative result. On the two suites where the AR backbone is strong, the backbones are indistinguishable: spatial (C 93.0 [90.4, 94.9] vs. D 94.2 [91.8, 95.9]) and goal (C 73.2 [69.2, 76.9] vs. D 71.4 [67.3, 75.2]) both have overlapping 95% Wilson intervals (p=0.44, p=0.52). On the two suites where the AR backbone is weak, the diffusion prior wins decisively and the intervals are disjoint: object (C 93.2 vs. D 82.0, +11.2, p<10⁻⁷) and long-horizon (C 69.4 vs. D 57.4, +12.0, p<10⁻⁴). This matches the hypothesis: D (and A′) degrade sharply on long, multi-stage tasks (A′−A is −13.5 on long), and it is exactly there that the bidirectional, diffusion-trained representation recovers the most. Because the comparison is at matched scale (7B), matched initialization, and matched vision/expert/recipe/seed, the gap is attributable to the pretraining that distinguishes the two backbones rather than to capacity or architecture. The significance is at the rollout level and states only that these two trained models differ; it does not speak to training-seed variance, which we could not measure at n=1 (§6.3, §7).

6.3Representation-level probing

The rollout comparison tells us whether the diffusion prior helps end-to-end; probing asks why, by measuring how much action-relevant information is already linearly decodable from the backbone's action-token hidden states before the flow expert acts. Because C and D share architecture, width (3584), vision, LoRA recipe, and expert, and differ chiefly in pretraining, a probe gap between them at matched capacity is attributable to the pretraining prior and mirrors the C−D contrast at the representation level.

Protocol. For each trained backbone we run a single frozen forward pass (k=1) over 4000 LIBERO observations at flow time τ=1 (pure-noise action input), and read the last-layer hidden states at the H=10 action-token positions, mean-pooled to a 3584-d feature. C and D are extracted in the same dataset order, so they are probed on identical frames; because both are 3584-d, one probe architecture is capacity-matched across them. We fit a linear probe (ridge for regression, logistic for classification) as our headline measure, and a small 2-layer MLP as a nonlinear ceiling, over 5 train/test splits.

Targets. We decode the action-relevant quantities available directly from the dataset: the immediate action and the full action chunk (7-DoF, R²), the proprioceptive state (R²), and the gripper command (median-split, accuracy). More abstract targets (subtask phase, target-object pose) require simulator-state annotation and are left to future work.

The gain is not a low-level decodability effect. At the representation level, C and D are tied: their action-token features are equally, and very highly, linearly decodable for every low-level target (|C−D|≤0.01 throughout), whereas the rollout gap is +5.9 on average and +12.0 on long-horizon. This cleanly rules out the simplest explanation of the behavioral gain: it is not that the diffusion prior makes the immediate action, state, or gripper command more linearly readable, since that information is present to the same (high) degree in both backbones. We do not overread the null: the probe does not localize the advantage, because both backbones sit near ceiling (0.87–0.96), the targets are low-level and read from single in-distribution frames while the rollout gap is a long-horizon closed-loop effect, and the nonlinear MLP probe was inconclusive. Higher-level or temporal structure and closed-loop robustness remain plausible mechanisms.

7Discussion

The matched C−D contrast is positive: the diffusion-trained backbone yields a better policy, and the advantage is largest on long-horizon control (+5.9 mean, +12.0 on long-horizon, both significant at the rollout level). Because C and D are identical up to the pretraining objective (same expert, vision, LoRA recipe, seed, and data), we read this as evidence that a bidirectional/denoising pretraining prior produces hidden states that a fixed flow-matching expert can exploit better than those of a matched autoregressive backbone, most in the long-horizon regime where the AR backbone degrades. The probing analysis sharpens the claim by ruling out the simplest account: if the diffusion prior simply made low-level control more linearly readable, C would out-probe D; instead the two are tied at ceiling on every low-level target while C clearly out-rolls D. Whether the effect is higher-level or temporal structure, or closed-loop robustness to compounding error, is left open for deferred abstract-target and closed-loop probes.

Two points of scope. In absolute terms C sits below A′ and A, but this reflects the decoupled-expert and LoRA-vs-full-fine-tuning costs, which are shared by C and D and cancel in the identifying C−D contrast. And while our design controls attention topology by giving C and D the same mask, a natural follow-up is a backbone with an expanded attention topology over the action block (a "Model B" built on D's weights), which would separate any benefit of richer action-block attention from the pretraining prior.

7.1Limitations

Training-seed variance. C and D are single-seed (n=1). The Wilson intervals and tests establish that the two trained models differ at the rollout level, but do not capture variance across retraining. Across seeds, A′ long-horizon success ranged 74.6–82.4 (sd ≈ 4.3), so single-seed suite gaps should be read as provisional. That the two significant gaps (+11.2, +12.0) are several times this seed spread is encouraging, but confirming them with additional seeds is the most important next step.

Objective vs. corpus confound. Dream-7B is initialized from Qwen2.5-7B and then trained with a discrete-diffusion objective, but that training also uses its own data and compute. With off-the-shelf backbones we cannot separate the pretraining objective from the corpus/compute that accompanied it; "pretraining prior" should be read as "the pretraining that produced Dream rather than Qwen," not the diffusion objective in isolation.

What is not confounded. Attention topology is controlled: C and D use the identical block-diagonal prefix-LM mask at training and inference, so the contrast is not explained by bidirectional-vs-causal attention usage. Scale, initialization, vision encoder, action expert, adaptation recipe, data, and seed are likewise matched.

Scope. Evaluation is simulation-only on a single benchmark (LIBERO), so real-robot external validity is untested. Absolute performance reflects the standalone (decoupled) expert, which costs ~6–7 points vs. the fused expert; this cost is shared by C and D and cancels in C−D. The probes target low-level quantities that saturate at ceiling for both backbones, so the representation-level comparison is suggestive rather than conclusive.

8Conclusion

We asked whether the pretraining objective of a VLA backbone, diffusion versus autoregressive, matters for downstream manipulation when everything else is held fixed, and built a controlled test to answer it: a single standalone flow-matching action expert driven by interchangeable 7B backbones that differ only in how they were pretrained (Dream, initialized from Qwen2.5-7B, versus Qwen2.5-7B itself). The diffusion-trained backbone is the better policy, improving mean LIBERO success by +5.9 points and by +12.0 on the long-horizon suite, the regime where the autoregressive backbone degrades most. Because the two systems are matched in scale, initialization, vision, action expert, adaptation recipe, data, and seed, we attribute the gap to the pretraining prior. At the same time, linear probing shows the two backbones' action-token representations to be equally decodable for low-level action, state, and gripper targets: the behavioral advantage does not reduce to the linear readability of immediate control signals. The study is single-seed, simulation-only, and probes a deliberately low-level target set, so we read it as a controlled existence proof rather than a final measurement: a diffusion-trained backbone can yield a better flow-matching policy than a matched autoregressive one, most where it matters, and not for the reason one would first guess.

 References

1. C. Chi et al. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. RSS 2023. arXiv:2303.04137.
2. Physical Intelligence. π0: A Vision-Language-Action Flow Model for General Robot Control. 2024.
3. K. Black et al. π0.5: A Vision-Language-Action Model with Open-World Generalization. CoRL 2025.
4. Physical Intelligence. OpenPI. github.com/Physical-Intelligence/openpi.
5. J. Ye, Z. Xie et al. Dream 7B: Diffusion Large Language Models. 2025. arXiv:2508.15487.
6. Qwen Team. Qwen2.5 Technical Report. 2024. arXiv:2412.15115.
7. S. Li et al. LaViDa: A Large Diffusion Language Model for Multimodal Understanding. 2025. arXiv:2505.16839.
8. DreamLM. Dream-VL / Dream-VLA. 2025–2026.
9. Y. Wen et al. LLaDA-VLA: Vision Language Diffusion Action Models. 2025.
10. Discrete Diffusion VLA. 2025. arXiv:2508.20072.
11. Google. DiffusionGemma. 2026.
12. X. Li et al. Crossway Diffusion. ICRA 2024. arXiv:2307.01849.
13. Y. Ze et al. 3D Diffusion Policy. 2024. arXiv:2403.03954.
14. C. Liu et al. Ordered Action Tokens. 2026. arXiv:2607.21670.
15. B. Liu et al. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. NeurIPS D&B 2023. arXiv:2306.03310.