acceptodds
Under review as a conference paper at ICLR 2027

Do We Really Need Multi-Step Refinement for VLA Planners?

Abstract

Pretrained vision-language models bring rich semantic understanding and trajectory reasoning to autonomous driving, spurring a growing family of Vision-Language-Action (VLA) planners. Existing approaches either adopt iterative decoding paradigms inherited from text and image generation, or employ downstream refinement modules widely used in robotic manipulation. However, neither is compatible with the stringent latency of autonomous driving. Since future trajectories are low-dimensional, kinematically constrained, and orders of magnitude simpler than images or text sequences, we challenge the need for such complex pipelines in VLA planning. To this end, we present SnapVLA, a framework that generates the complete trajectory in a single parallel forward pass via masked discrete diffusion. We demonstrate that the perceived demand for multi-step refinement originates from two factors that do not inherently require it: the policy optimization paradigm and the local incoherence among trajectory points. Regarding optimization, single-step diffusion yields an exact sequence likelihood and an unbiased GRPO objective, whereas existing multi-step chains admit no exact likelihood. Regarding coherence, we tokenize trajectories into a structured action space and apply a lightweight closed-form smoother, yielding kinematically consistent trajectories without a learned continuous expert. On NAVSIM v1 and v2, SnapVLA achieves 91.0 PDMS and 90.1 EPDMS, on par with state-of-the-art models. Notably, it takes 180 ms per inference on a single A800 GPU, over 2 faster than the fastest baseline, demonstrating the potential of the single-step paradigm for efficient planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, autonomous driving

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_y9koOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SnapVLA addresses the efficiency of action generation in VLA models and achieves both strong performance and higher efficiency. The method is very promising. Experiments on more benchmarks would further support the paper’s claimed contributions.

Critical strengths
  1. Efficiency is a key issue for VLA models. The paper achieves strong performance while maintaining efficient action generation.

  2. The ablation studies are fairly complete.

  3. The writing and figures are easy to follow.

Critical weaknesses
  1. The paper focuses on the advantages of one-step generation. A performance–latency curve comparing 1, 2, 4, and 12 generation steps, using the same binned representation, training budget, and smoother, would better support this claim. It could also show the method’s potential for multi-step generation.

  2. NAVSIM v1 and v2 essentially evaluate different metrics on navtest. The method should also be tested on other datasets to assess its generalization, such as navhard, which emphasizes challenging scenarios, and the Waymo end-to-end dataset. Also, the NAVSIM benchmark is based on nuPlan, rather than nuScenes.

  3. The paper only studies VLA action generation for autonomous driving. It does not explore manipulation or robot navigation, so the title should explicitly mention autonomous driving. The current title appears to claim a broader scope than the experiments support.

Questions for authors

Please refer to the weaknesses section.

Reviewer_xLm5OpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether VLA planners need multi-step decoding or an extra continuous action expert, and argues that they do not. SnapVLA is built on LaViDa-8B with a SigLIP-so400M visual encoder. Trajectories are represented as binned discrete tokens. The longitudinal and lateral coordinates over the range [-100, 100] m are quantized into 4000 bins at 0.05 m resolution. The heading over [-1.7, 1.7] rad is quantized into 340 bins at 0.01 rad. Each waypoint uses three tokens, so a trajectory with K waypoints has L = 3K tokens. Training uses masked discrete diffusion: tokens are replaced by [MASK] with probability r, and the model reconstructs them. At inference every position is masked and all tokens are predicted in one parallel forward pass. The paper gives two reasons why single-step decoding works. First, one-step masked diffusion gives an exact sequence log-likelihood, because all tokens are sampled independently from one fully masked input. A multi-step chain requires marginalizing over all intermediate denoising states, and the number of configurations grows exponentially in both sequence length and step count, so existing methods use approximate likelihoods. The exact likelihood makes the GRPO importance ratio exact. Second, local incoherence between independently predicted waypoints is handled by a closed-form smoother that minimizes a third-order finite-difference penalty with fixed endpoints, solved as a linear system, with lambda set to 0.05. Training has four stages: VLM adaptation on 800k VQA samples and 400k trajectory samples, SFT on NAVSIM for 5 epochs, GRPO for 2 epochs with G = 16 and a reward built from PDMS components, then the smoother. Reported results are 91.0 PDMS on NAVSIM v1 and 90.1 EPDMS on NAVSIM v2, at 0.18 s per inference on a single A800 GPU.

Critical strengths

The exact-likelihood argument is the part I value most. The paper shows that single-step decoding is faster, and it is also the cleaner setting for policy optimization. With all tokens sampled from one fully masked input, the sequence log-likelihood is the sum of the per-position log probabilities and can be computed exactly. With S denoising steps, computing the probability of the final trajectory requires summing over intermediate token configurations, and that number grows exponentially in both L and S. Existing masked-diffusion planners therefore rely on path-based or factorized approximations. SnapVLA does not need those approximations. The efficiency argument also becomes a correctness argument here, and I find that more convincing than a latency table on its own.

The second strength is the latency result, and it is measured well. Table 4 reports every number on one NVIDIA A800 GPU, which makes the numbers comparable with the rest of the literature. The most relevant comparison is against DriveFine, which uses the same LaViDa-8B backbone and the same masked-diffusion family. SnapVLA reports 91.0 PDMS at 0.18 s against 90.7 PDMS at 0.39 s. That is a real speedup at matched accuracy, on the same hardware and with the same backbone.

The binned representation result in Table 1 is also clean evidence. Unbinned tokens give 66.4 percent valid decodings in one step. Binned tokens give 100 percent. Table 5 supports the same point: unbinned one-step decoding gives 57.5 PDMS, and binned one-step decoding gives 88.9 PDMS.

Critical weaknesses

My main concern is the smoother. In Table 5, adding the smoother changes EPDMS on NAVSIM v2 by 4.1 points (86.0 becomes 90.1), and EC by 39.9 points (42.8 becomes 82.7). The change on NAVSIM v1 is zero, because PDMS uses the comfort submetric C and SnapVLA already reports 100 on C. So the entire NAVSIM v2 result, which is the highest value in Table 3 and the basis for the state-of-the-art claim, comes from a deterministic post-processing filter. Without it, 86.0 EPDMS would place the model below DriveFine and ExploreVLA. The filter itself is reasonable, and Table 6 shows that jerk drops by roughly a factor of 14 (3.632 becomes 0.254). Two things still bother me. Lambda is fixed at 0.05 with no sweep reported, and the paper does not say which split lambda was chosen on. If lambda was selected on navtest, part of the 4.1 point gain is test-set tuning. The second thing is that EC is largely a jerk-based metric, and the smoother explicitly minimizes third-order differences. I would like the full submetric breakdown with and without the smoother. Does NC, TTC, or EP change at all, or does the entire gain come from EC?

The second weakness is a missing control. The paper claims that multi-step refinement is unnecessary. The evidence given is that unbinned decoding needs many steps and binned decoding does not. The control I want is the same binned model run with 2, 4, 12, and 48 steps. If binned multi-step decoding beats binned one-step decoding, then the honest claim is a speed-accuracy trade-off, and the question in the title has a different answer. If binned multi-step decoding is no better, the claim is much stronger. This experiment does not need new training, because the model is already trained with a masking-ratio distribution, and I think it is the most important missing result.

Third, there is no variance estimate. NAVSIM navtest has 136 scenes. The margins over the strongest baselines are 0.2 PDMS on v1 and 0.4 EPDMS on v2. GRPO also adds sampling stochasticity. One run without error bars on 136 scenes is not enough for me to rank these methods.

Smaller points. Table 4 gives total latency with no breakdown. With an 8B backbone I would like to know how much of the 0.18 s is the vision encoder and the VLM prefill and how much is the diffusion head, since that determines how much a faster backbone would help. The GRPO stage uses a filtered NAVSIM subset and the filtering rule is not described. The vocabulary change is not discussed: 4000 plus 340 action tokens is a large addition, and most of the longitudinal bins beyond roughly 60 m are probably never used. The model uses monocular front-view input only, which the paper states as a limitation. There is no navhard result, only navtest, and a reactive evaluation would strengthen a paper whose main argument is about deployment.

Questions for authors

Can you report a lambda sweep, and state which split lambda = 0.05 was selected on?

Can you show all NAVSIM v2 submetrics with and without the smoother? I want to know whether the EPDMS gain comes entirely from EC.

Can you run the binned model with 2, 4, 12, and 48 denoising steps and report both the metrics and the latency?

Can you report error bars or multiple seeds on navtest, given the 0.2 and 0.4 margins?

What is the latency breakdown between the vision encoder, the VLM prefill, and the diffusion head?

How was the GRPO subset filtered, and is it drawn from navtrain?

How many action tokens are added to the vocabulary, and what fraction of the 4000 longitudinal bins are ever used?

Reviewer_s3wdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed a VLA planner built on LaViDa-8B, a masked-diffusion VLM. It predicts the whole ego trajectory in a single forward pass from a fully masked token sequence. Waypoints are discretized into coordinate-specific bins: 4000 bins over ±100 m for xx and yy, and 340 bins for heading. This binned representation follows DriveFine.

Training has three learned stages: Driving adaptation on ~800k VQA and ~400k trajectory samples, SFT on NAVSIM with a masked-token reconstruction loss, and a GRPO with a reward built from PDMS components.

For GRPO, the sequence likelihood is the product of per-token marginals given the fully masked input. The authors present this as exact and unbiased, in contrast to the approximate likelihoods used for multi-step masked-diffusion planners. At inference, the de-quantized trajectory goes through a closed-form ridge smoother that penalizes third-order differences, with both endpoints fixed.

Reported results are 91.0 PDMS on NAVSIM v1 and 90.1 EPDMS on v2, at 0.18 s per plan on an A800. The fastest baselines take 0.36 to 0.39 s. The ablation (Table 5) shows that: Binning is what makes one-step decoding work (57.5 → 88.9 PDMS), the GRPO adds 2.1 PDMS but lowers EC (49.8 → 42.8), and the smoother leaves PDMS unchanged and raises EPDMS from 86.0 to 90.1, almost entirely through EC (42.8 → 82.7).

Critical strengths
  1. The practical result is useful if it holds. An 8B VLM planner reaches 91.0 PDMS from a single camera at roughly half the latency of DriveFine, which uses the same LaViDa-8B backbone and so makes the most controlled point in Table 4. For people trying to deploy VLM planners, a working single-pass recipe with no action expert is worth knowing about. This is what keeps me from a clear reject.

  2. The staged ablation (Tables 5 and 6) makes clear where each gain comes from. It also reports the EC drop under GRPO instead of hiding it. That makes the paper easy to argue with, and several points in W1 come directly from this table.

Critical weaknesses
  1. The paper’s central question (“do VLA planners need multi-step refinement?”) is not answered by the experiments, and the paper’s own ablation points the other way on v2.
  • The key control is missing, and the binned model is only ever decoded in one step. To claim multi-step refinement is unnecessary, the same binned model (after SFT and after GRPO) needs to be decoded at 2, 4, 8 and 12 steps with latency reported. Table 1 and Table 5 only run multi-step decoding for the unbinned representation.
  • In Table 5, the unbinned 192-step model (row 3: 87.9 EPDMS, 87.1 EC) beats the binned single-step model both after SFT (85.5 / 49.8) and after GRPO (86.0 / 42.8). SnapVLA only leads on v2 after a hand-designed post-processing step. That step is refinement in all but name. Even after smoothing, EC (82.7) is below that unbinned multi-step model (87.1). It is also below DriveFine (83.8), ReCogDrive (86.5), ExploreVLA (86.8) and SpanVLA (85.1). So “single-step generation is sufficient” holds for v1 PDMS, which the smoother doesn’t change. For v2 it is better read as “single-step generation plus a smoother.”
  • The DriveFine comparison in Fig. 1(b) isn’t fair. DriveFine was trained for 12 steps and then evaluated at one step. The fair version would apply the same smoother to baseline outputs (at least DriveFine and row 3 of Table 5). If the smoother lifts them by a similar amount, the v2 advantage has nothing to do with the single-step policy.
  • The v1 margin is 0.2 to 0.4 PDMS over DriveFine, ReCogDrive, ResAD and AutoDrive-P3, from a single run with no variance reported.
  • The tables disagree with each other. ReCogDrive is listed at 90.8 PDMS in Table 2 and 89.6 in Table 4.
  • The speedup claim is overstated. “Over 2× faster than the fastest baseline” is exactly 2.0× against ReCogDrive (0.36 vs 0.18 s). The latency protocol is also not described: precision, KV caching, step counts per baseline, and whether vision encoding is included.
  • The binned tokens come from DriveFine, and third-order-difference ridge smoothing is a classical Whittaker-type smoother. What’s left is the combination and the empirical claim, which makes the missing controls more important.
  • The closest prior work is missing from the comparison. WAM-Diff, a masked-diffusion VLA with online RL, is cited but absent from Tables 2 to 4, while WAM-Flow from the same group is included.
  1. The exact-likelihood argument is weaker than presented, and it is never tested.
  • From a fully masked input in one step, the policy is a product of independent 4000-way categoricals. Nothing models the correlation between xx, yy and ψ\psi of one waypoint, or between neighboring waypoints. The paper replaces likelihood approximation with model misspecification and doesn’t discuss the trade. Independent per-token sampling during GRPO will mix modes and add jitter. That is consistent with EC falling after GRPO and with needing a smoother afterwards.
  • The marginal likelihood of z0z_0 under a multi-step sampler is intractable, as the paper says. But policy-gradient methods don’t need it. If the denoising chain is treated as an MDP, the path likelihood is exactly computable. The resulting gradient is unbiased for the expected reward of the sampler that is actually deployed. With confidence-based unmasking, the path probability is just a product of chosen-token probabilities. So the claim that path-based estimates create a gap between optimization and deployment needs an argument the paper doesn’t make.
  • GRPO rewards are computed on raw de-quantized trajectories, but deployment applies the smoother afterwards. If deployment also uses argmax rather than sampling, that is a second mismatch. Either way, “aligns optimization with the deployed policy” doesn’t hold as stated.
  • No experiment compares exact and approximate likelihoods on the same model. For example, the paper could train the single-step policy with a mean-field or path-based estimator, or train a binned multi-step policy with a path-likelihood ratio. Without this, the abstract’s “we demonstrate that the demand for multi-step refinement originates from … the policy optimization paradigm” is not supported.
Questions for authors
  1. Is deployment greedy or sampled? What sampling temperature is used during GRPO? Are rewards computed on raw or smoothed trajectories?

  2. Which ReCogDrive number (Tables 2 vs 4) is correct? Are the Table 4 numbers reproduced by the authors?

  3. For each method in Table 4, give precision, batch size, KV caching and the number of decoding steps. For SnapVLA, give a per-stage breakdown (vision encoder, prefill, decoding, smoother).

  4. What are the sources of the 800k VQA and 400k trajectory samples in Stage 1? Please confirm they don’t overlap with navtest logs. What were the filtering criteria for the GRPO subset?

  5. How was λ=0.05\lambda = 0.05 chosen? If it was tuned on navtest, the 86.0 → 90.1 EPDMS gain is tuned on test. How sensitive are EPDMS and EC to λ\lambda? Variance. What is the spread across seeds for 91.0 PDMS and 90.1 EPDMS?

  6. Please run 2/4/8/12 steps after SFT and after GRPO, with latency. This is the experiment that is needed to answer the paper’s central question (“do VLA planners need multi-step refinement?”).

  7. Should we apply the same smoother to DriveFine (1 and 12 steps) and to Table 5 row 3.

  8. Please compare exact vs approximate likelihood on the same backbone.

  9. Is “masked diffusion” needed? Inference always starts from r=1r=1 with one step, so the deployed model is a parallel per-token classifier. Does training only at r=1r=1 match the reported results? Does a plain regression head on the same LaViDa features match them? If either does, the masked-diffusion framing isn’t what makes single-step decoding work.

  10. NAVSIM is built on OpenScene/nuPlan, not nuScenes. The 1,192 / 136 figures are logs, not scenes; navtrain and navtest contain roughly 100k and 12k samples.

  11. Fig. 3 shows four camera views, while Section 4.4 says only the monocular front view is used. Please clarify.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.