acceptodds
Under review as a conference paper at ICLR 2027

Hybrid-Flow Drive: Unifying Masked Diffusion and Continuous Embedding Flow for Autonomous Driving

Abstract

End-to-end autonomous driving with vision-language-action (VLA) models requires trajectories that are accurate, consistent with the observed scene and navigation command, and reliable in long-tail cases. Diffusion language models provide an efficient alternative to purely autoregressive generation, but standard discrete masked-diffusion decoding is not well matched to planning: a fixed [MASK] token carries little information about the current noise level, supervision is often limited to corrupted positions, and standard decoding repeatedly discards partial evidence at unresolved tokens. We present Hybrid-Flow Drive, a hybrid flow diffusion VLA that replaces static masks with continuous, noise-aware token embeddings. Our approach has three main innovations: (1) Hybrid Flow Language Model (HFLM), which corrupts response embeddings in continuous space and gives each unresolved position an evolving state; (2) Next-Token-Invisible Bidirectional Attention (NBA), which prevents ground-truth leakage under continuous embeddings and allows dense supervision over response tokens; and (3) Verified Flow Decoding (VFD), which keeps the continuous state across denoising steps and checks the drafted sequence with a causal branch to recover autoregressive consistency. On the Waymo Open Dataset for End-to-End Driving, Hybrid-Flow Drive achieves a highly competitive Rater Feedback Score of 7.772, surpassing prior trajectory-supervised baselines. On nuScenes, it achieves an average L2 trajectory error of 0.31 m. Our results demonstrate that continuous noise-aware states can improve the human-aligned planning quality of diffusion-style end-to-end autonomous driving.

Primary area:
applications to robotics, autonomy, planning
Keywords:
end-to-end autonomous driving, vision-language-action model, masked diffusion, flow matching, speculative decoding

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 4 reviews, above 27% of the others.
Reviewer_mG87OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

is work presents Hybrid‑Flow Drive, a hybrid flow‑diffusion vision‑language‑action model for end‑to‑end autonomous driving. It replaces static discrete [MASK] tokens with noise‑aware continuous embeddings via the Hybrid‑Flow Language Model (HFLM). The paper further proposes Next‑Token‑Invisible Bidirectional Attention (NBA) for dense leakage‑free supervision and Verified Flow Decoding (VFD) to maintain continuous intermediate states with autoregressive draft verification.

Critical strengths

The hybrid design between masked diffusion and continuous embedding flow is well‑motivated for driving VLA planning. It keeps discrete token prediction while encoding noise magnitude into unresolved token embeddings, avoiding drawbacks of both static mask diffusion and fully‑continuous language flow models.

Critical weaknesses
  1. All evaluations are open‑loop only. Real‑world autonomous driving requires closed‑loop simulation validation; open‑loop metrics cannot fully reflect driving safety and agent interaction performance. 2.There is insufficient analysis regarding computational overhead. Although token throughput improves, the per‑step transformer cost of the dual‑stream V‑NBA inference is not quantitatively analyzed.
  2. The hyperparameter sensitivity for key components (denoising step N, confidence threshold τ, β loss weight for V‑NBA) lacks systematic ablation, making it unclear how robust the performance gain is across hyper‑configurations.
Questions for authors

Provide additional experimental analysis to address the identified weaknesses.

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

This paper presents Hybrid-Flow Drive, a vision-language-action (VLA) approach for end-to-end autonomous driving that combines continuous embedding-space corruption with discrete token reconstruction and autoregressive verification. The paper is motivated by limitations of masked diffusion language models for planning: a shared [MASK] embedding does not encode the degree of uncertainty, supervision is sparse because it is limited to masked positions, and iterative masked decoding does not retain partial evidence across steps.

The proposed method has three components:

Hybrid Flow Language Model (HFLM): replaces static masks at editable response positions with continuous, noise-corrupted token embeddings interpolating between Gaussian noise and ground-truth embeddings, while retaining cross-entropy reconstruction of discrete trajectory tokens. Next-Token-Invisible Bidirectional Attention (NBA): enables dense supervision of response tokens while masking each target from its predictor to avoid direct target leakage. Verified Flow Decoding (VFD): iteratively transports unresolved embedding states toward predicted clean embeddings, freezes high-confidence tokens, and uses an autoregressive branch to verify the accepted trajectory-token prefix.

Critical strengths

The paper identifies concrete drawbacks of static masked diffusion for structured driving outputs: uniform unresolved-token representations, sparse reconstruction supervision, and loss of accumulated evidence between iterative refinement steps. This is a coherent motivation for introducing a stateful continuous representation.

HFLM does not attempt to make all aspects of language/action modeling continuous. Instead, it uses continuous embeddings only for unresolved response states while retaining discrete vocabulary prediction and cross-entropy reconstruction. This is a practical design choice for structured trajectory-token generation, where token fidelity matters.

The next-token-invisible attention pattern is a thoughtful attempt to combine bidirectional response context with dense, AR-style supervision while avoiding direct access to the target token. The incremental ablation suggests NBA contributes meaningfully beyond HFLM alone.

Maintaining continuous states for undecided positions, committing confident tokens, and verifying the resulting draft with an autoregressive branch is a reasonable way to avoid repeatedly reverting to identical mask representations. The design also gives a clear explanation for why the method can commit multiple tokens per decoding step.

Table 2 is more convincing than the broad external comparison because it controls the backbone, data format, and training recipe across AR, masked-diffusion, block-diffusion, and Hybrid-Flow Drive variants. Under this setting, the method improves both RFS/ADE and reported decoding throughput.

The ablation on HFLM, NBA, and VFD is concise but informative. Relative to HFLM alone, adding both NBA and VFD improves ADE@3s from 0.6692 to 0.4604 and ADE@5s from 1.8096 to 1.4223 on the WOD-E2E validation subset.

The authors explicitly state that evaluation is open-loop and that the nuScenes setting is restricted to three front-camera frames. This transparency is appreciated.

Critical weaknesses
  1. The central methodological framing is a hybrid flow diffusion VLA. However, HFLM is trained using a fixed continuous corruption process and discrete cross-entropy reconstruction. During inference, VFD uses an analytic straight-line transport toward the current clean-token estimate, followed by Euler updates. The manuscript explicitly states that this is “rather than a learned velocity-matching objective.”

Thus, the paper does not appear to learn a continuous vector field or optimize a conventional flow-matching objective. The core training objective is still token reconstruction under continuous embedding corruption. This may be a useful method, but the terminology risks overstating the relationship to flow matching.

The paper should more clearly distinguish:

continuous embedding corruption, analytic interpolation/transport at inference, diffusion-style iterative decoding, and learned flow matching or continuous normalizing flows. At minimum, the claims should be calibrated. A more precise description might be “continuous-state masked decoding with AR verification,” unless the authors provide an argument showing that their objective is formally equivalent to, or a valid estimator of, a flow-matching objective.

  1. The claimed efficiency is a central contribution: 335.30 tokens/s compared with 45.73 for AR and approximately 100 for masked/block diffusion. However, tokens/s and tokens/step alone are insufficient to evaluate deployment relevance for driving.

In particular, VFD includes:

one or more HFLM drafting passes, multiple Euler steps, confidence-based commitment, an AR verification pass, potential corrective decoding after a mismatch. The paper should report end-to-end latency per planning query—including image encoding, all drafting steps, AR verification, and post-processing—under a fixed hardware setup and fixed output length. It should also report latency distribution, e.g., mean and p95, since the number of verification rounds likely varies per scene. Without this, it is difficult to assess whether the speedup is realized in a realistic driving stack rather than only in an effective-token-throughput abstraction.

A helpful comparison would include:

wall-clock latency per sample, trajectory tokens generated per sample, number of HFLM and AR forward passes, number of VFD outer verification loops, GPU memory, and accuracy–latency trade-offs as the number of ODE steps and confidence threshold vary.

  1. The method is positioned as improving planning quality, global consistency, and reliability in long-tail driving scenarios. Yet the primary evaluation is open-loop: WOD-E2E uses RFS and ADE, and nuScenes uses trajectory L2. These are useful but do not evaluate compounding errors, interactive behavior, traffic-rule compliance, or safety under the policy’s induced state distribution.

This is especially important because VFD’s claimed benefit is globally consistent iterative trajectory generation. Such a claim is best supported by closed-loop evaluation, where a trajectory policy must react to its own prior behavior. The paper should include a closed-loop benchmark—e.g., CARLA/Bench2Drive or an equivalent simulator—with route completion and safety metrics such as collision rate, red-light violations, off-route events, and comfort.

If closed-loop evaluation is unavailable, the claims should be narrowed to open-loop trajectory prediction and human-aligned trajectory scoring, rather than robust end-to-end planning.

  1. The paper acknowledges that Table 1 aggregates methods with different output formats, training recipes, data, supervision, and possibly sensor inputs. This caveat is correct, but it substantially limits the table’s ability to establish state-of-the-art performance. In particular, some methods may use CoT, different perception supervision, different camera configurations, distinct preprocessing, or different model scales.

The controlled comparison in Table 2 is much stronger evidence. The paper should elevate it to the main empirical claim and avoid presenting the cross-paper comparison as decisive. For nuScenes, the authors appropriately note that their method uses three front-camera frames and that compared methods may differ in sensors, histories, supervision, and scale; the same level of caution should apply to the WOD-E2E comparison.

  1. The ablation establishes that HFLM helps over the AR baseline, but it does not isolate why. HFLM changes several things at once: embedding-space corruption, continuous time/noise level, corruption schedule, the decoding process, and perhaps optimization behavior. Therefore, the central claim—that evolving continuous states are superior to a shared static mask representation—is not fully demonstrated.

I would like to see stronger controlled studies such as:

Static-mask versus continuous embedding corruption under identical NBA/VFD or identical iterative decoding; Continuous embeddings with no retained state between steps versus VFD state retention; a discrete masked-diffusion baseline with dense supervision, if feasible; the effect of γ, corruption-time distribution, and noise schedule; quality and calibration of confidence scores used for freezing; performance as a function of number of Euler steps and verification rounds. These experiments would clarify whether the gains come from continuous uncertainty representation specifically, or primarily from dense supervision and AR verification.

  1. NBA is critical: it enables dense supervision precisely where naïve bidirectional attention would leak the target embedding. Although the paper provides a visibility-mask formulation and Appendix A describes V-NBA, the construction remains difficult to verify from the presentation alone—particularly given the simultaneous clean and corrupted response streams and shared positional identifiers.

A stronger empirical validation would be valuable:

train/evaluate a deliberately leaky bidirectional variant and show its misleadingly low training loss but poor generation; provide a small toy demonstration that each target is inaccessible from its predictor; state precisely which response embeddings are visible to each predictor position; clarify whether target information can be inferred indirectly through the clean stream or shared positional alignment. Because the proposed dense objective is a principal contribution, this detail should be unambiguous and reproducible.

  1. Robustness and statistical reliability are underexplored The reported improvements, particularly on nuScenes, are small relative to the best listed average L2 values: 0.31 versus 0.33 m. Moreover, Hybrid-Flow Drive has a slightly worse 3-second error than the best listed prior method (0.57 versus 0.55 m). It would be helpful to report results across multiple random seeds or confidence intervals, especially for the controlled WOD-E2E comparisons and the small cross-dataset difference.

Further, claims about long-tail reliability would be more convincing with breakdowns by:

scenario type or difficulty, turning/merging/intersection cases, adverse visibility or agent density, horizon, rare safety-critical events, and navigation-command categories.

Questions for authors

Why is this method best characterized as flow matching rather than continuous embedding corruption plus analytic denoising toward predicted tokens? Is there a formal connection between the proposed objective and a flow-matching objective?

What is the end-to-end latency per driving decision, including visual encoding, all HFLM iterations, AR verification, and trajectory decoding? Please report mean/p95 latency and memory for comparable AR, MDM, block-diffusion, and Hybrid-Flow Drive systems.

How many VFD outer-loop verification rounds occur on average, and what fraction of drafted tokens are accepted by the AR verifier? How does this vary by scene complexity?

Can the authors isolate the contribution of continuous unresolved states from dense NBA supervision and AR verification using more tightly controlled ablations?

Does NBA/V-NBA provably prevent direct or indirect target leakage? Can the authors provide either a formal argument or a diagnostic experiment with a deliberately leaky control?

Why is nuScenes evaluated with front-only temporal input while WOD-E2E uses three surround cameras? Is there a WOD-E2E-like multi-view nuScenes protocol or a front-camera matched baseline comparison?

Can the authors provide closed-loop evaluation, or otherwise restrict claims about planning robustness and safety to the open-loop setting?

Are the results statistically stable across seeds? This is particularly relevant for the relatively small nuScenes margin and the controlled WOD-E2E improvements.

Minor Comments The paper should report the trajectory-token serialization format, output length, coordinate convention, quantization resolution, and how token errors map to spatial trajectory error.

It would help to report whether HFLM/NBA/VFD improve performance uniformly across waypoint horizons or primarily at short horizons. The nuScenes result is better at 1s and 2s but not best at 3s.

Please provide details about the WOD-E2E split, the construction of training examples, and whether any scenes overlap semantically or geographically with evaluation conditions.

Table 1 should make sensor configuration, temporal context, model size, use of privileged supervision, and training data explicit for all methods where available.

The speed benchmark needs reproducibility details: precision, batch size, warm-up procedure, decoding length, use of KV cache, and whether image encoding is included.

The manuscript’s contribution would be clearer with an explicit algorithm box for training as well as the current inference algorithm.

The term “human-aligned” should be used carefully. RFS is derived from rater-annotated reference trajectories and trust regions, but this does not by itself establish human-aligned behavior beyond that metric.

Reviewer_amaROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Hybrid-Flow Drive, a hybrid flow diffusion VLA for autonomous driving. Specifically, it replaces static masks with continuous, noise-aware token embeddings, while NBA enables dense token supervision without target leakage.

In addtion, the proposed VFD (Verified Flow Decoding) combines iterative flow-based refinement with autoregressive verification, achieving a 7.33× decoding speedup over the AR baseline.

However, the backbone (Qwen2.5) used in this method is quite dated. Furthermore, there is a lack of comparison with more recent methods. Therefore, the conclusions reached in the paper may be biased.

Critical strengths
  1. The proposed HFLM (Hybrid Flow Language Model) replaces static mask tokens with continuous noise-aware embeddings, while NBA enables dense token supervision without target leakage.

  2. The proposed VFD (Verified Flow Decoding) combines iterative flow-based refinement with autoregressive verification, achieving a 7.33× decoding speedup over the AR baseline.

Critical weaknesses
  1. Outdated backbone and insufficient comparisons. The paper only uses Qwen2.5-VL. The authors should consider a newer backbone, such as Qwen3.5, and add comparisons with recent methods [2][3][4] in Tables 1 and 6. Additionally, inference latency should be reported under comparable hardware settings.

  2. Unclear novelty of NBA. Fast-dVLM [1] already introduces similar attention mechanisms and autoregressive verification. The authors should clarify the differences between NBA/V-NBA and Fast-dVLM [1], and provide experimental comparisons.

  3. Similarity to existing flow-matching methods. The proposed method shares similarities with previous flow-matching-based approaches [2][3][5]. What are the key differences between these methods? The authors should clarify these differences in the Introduction or Related Work section.

  4. Lack of closed-loop evaluation. The authors should provide additional experiments on closed-loop benchmarks, such as Bench2Drive or HUGSIM, to further demonstrate the effectiveness of the proposed method.

[1] Fast-dVLM: Efficient Block-Diffusion VLM via Direct Conversion from Autoregressive VLM

[2] WAM-Flow: Parallel Coarse-to-Fine Motion Planning via Discrete Flow Matching for Autonomous Driving

[3] MindVLA-U1: VLA Beats VA with Unified Streaming Architecture for Autonomous Driving

[4] Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving

[5] GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving

Questions for authors

Was Figure 1 generated by AI? Regardless, please refine Figure 1, as it appears to be AI-generated.

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

This paper studies trajectory generation for autonomous driving: given camera images, vehicle state, and a navigation command, it plans how the vehicle should drive next. The authors propose Hybrid-Flow Drive, which generates trajectory candidates in a continuous vector space and then checks the result through an autoregressive branch. On WOD-E2E, the authors report improvements in trajectory quality and generation speed, and they evaluate trajectory-prediction performance on nuScenes. Overall, this work combines parallel generation with autoregressive checking and explores a concrete approach to improving the efficiency of driving-trajectory generation.

Critical strengths
  1. The research goal is practically meaningful. The authors study how to generate driving trajectories faster, providing a concrete direction for reducing the inference cost of trajectory planning.
  2. It includes a valuable controlled comparison. Under the same training configuration and model scale, the system raises the trajectory score from 7.500 to 7.772 relative to the autoregressive baseline, while significantly improving decoding throughput.
Critical weaknesses
  1. NBA cannot guarantee the target invisibility as defined in the paper. When predicting the 2nd token, although position 1 cannot directly read position 2, position 3 can read position 2, and position 1 can in turn read position 3 in the next layer. Therefore, the target information can still be routed back to the prediction position, bypassing the mask, which directly undermines the "leakage-free all-position supervision" that the paper emphasizes.
  2. The main efficiency experiment is in tension with the advantage of continuous states that the authors emphasize. Fast-dVLM has already adopted the scheme of "generate candidates, then verify with an autoregressive branch." This paper emphasizes progressive refinement by retaining continuous intermediate states, but Table 2 updates only a single step, which by Eq. (5) directly collapses to the embedding of the predicted token. Hence, this set of results does not yet demonstrate the additional value of continuous states over existing draft-and-verify schemes.
Questions for authors

NBA claims that the prediction position cannot see the target token. However, by Eq. (6), when predicting the 2nd token, position 3 can first read position 2 and then pass the information to the prediction position 1 in the next layer.

In the actual implementation, is there any additional mechanism that blocks this indirect path?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.