acceptodds
Withdrawn from ICLR 2027

Discrete Forcing: Infusing Discrete Guidance into Continuous Denoising for Few-Step Action Experts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Efficient action generation in vision-language-action (VLA) models requires capturing both coarse action structure and fine-grained details. Discrete action tokens provide compact structural representations but sacrifice precision, while continuous action tokens offer high precision but often require multiple denoising steps. We introduce Discrete Forcing, a flow-matching framework that combines these representations through an explicit coarse-to-fine generation process. It first predicts discrete action tokens to establish a coarse action structure, then uses them to guide continuous action refinement. The discrete and continuous components share a common diffusion transformer backbone with specialized branches, maintaining a parameter count comparable to a conventional single-branch model while requiring only one forward pass per branch. Extensive evaluations across multiple benchmarks demonstrate improved performance and faster inference over a parameter-matched continuous action expert, with consistent performance gains as model capacity increases. Real-world experiments further demonstrate improvements on high-precision and dynamic manipulation tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action Model, Robot Manipulation

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_M2eLOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Discrete Forcing, a coarse-to-fine action-generation framework for vision-language-action models. It first predicts quantized discrete action tokens from masked inputs, then uses their dequantized values as a structured source for one-step continuous flow-matching refinement. The discrete and continuous branches share early Action DiT layers but use specialized later branches. Experiments on LIBERO, RoboTwin 2.0, RoboCasa-GR1, and four real-world manipulation tasks report improved success rates and lower latency than continuous-action baselines, including 97.6% on LIBERO with two function evaluations and a 16.4-point average improvement in the real-world task-progress metric.

Critical strengths

1.The discrete-to-continuous formulation is conceptually clear and technically motivated. The method directly targets the trade-off between compact coarse structure and precise continuous control, while the ablations on generation order, discrete guidance, one-step training, and shared layers support the proposed mechanism.

2.The empirical evaluation is broad and practically relevant. The paper reports results across single-arm, dual-arm, humanoid, long-horizon, and real-world manipulation settings, together with latency, memory, scaling, and trajectory-structure analyses. The gains at two function evaluations are consequential for deploying VLA policies under real-time constraints.

Critical weaknesses

1.The claimed parameter matching is not sufficiently verifiable. The implementation describes 2 shared layers and 22 branch-specific layers per execution path, so storing both branches could require substantially more parameters than a single 24-layer action expert unless widths or other components are adjusted. The paper should provide exact total stored parameters, trainable parameters, inference memory, and training compute for Discrete Forcing and every matched baseline. Without this, the performance gains may partly reflect additional capacity rather than the discrete-to-continuous design.

2.The causal evidence for the proposed forcing mechanism remains incomplete. During training, the continuous branch primarily receives ground-truth quantized actions, while inference uses predicted discrete actions; the predicted-source ablation is reported only on LIBERO-Long. In addition, most comparisons rely on StarVLA, and the real-world evaluation uses a π0-based instantiation with four hand-designed tasks and 20 trials per task. More evidence is needed to establish robustness to discrete prediction errors and to separate architectural benefits from backbone-, dataset-, or task-specific effects.

Questions for authors

1.What are the exact total parameter counts and stored-weight memory for Discrete Forcing versus StarVLA at each model scale, including both discrete and continuous branches? Are training FLOPs and optimizer states also matched?

2.How often do predicted discrete tokens differ from the ground-truth quantized source during training/evaluation, and do the predicted-source results hold across the other benchmarks and model scales? Please clarify whether any ground-truth discrete information is available at inference.

3.For the simulation and real-world comparisons, were all baselines trained on identical demonstrations, optimization budgets, and evaluation seeds? Please provide per-task results and uncertainty estimates for the real-world task-progress numbers.

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

The paper presents a new method, namely Discrete Forcing (DF). The main idea of the method is a coarse-to-fine approach for action generation. Specifically, discrete action tokens are first predicted to get the overall structure of an action trajectory, and then used that for guiding a continuous refinement.

The core component of the proposed network is a shared Action DiT, with two branches (discrete and continuous). The discrete branch predicts action bins from a fully masked action sequence. Then the continuous branch performs a flow-matching refinement step.

Extensive experimental results were presented, and the proposed method achieves a higher success rate while having lower latency.

Critical strengths
  • Simple yet effective: the core idea of the proposed method is simple in general - with discrete token prediction and then refining the initial action trajectory structure through the continuous branch. The simple design also does not require iterative discrete decoding or multi-step continous denoising, and results in lower latency.

  • Extensive experiments: both simulation benchmarks and real-world evaluation were presented, and the improvement over other baselines are consistent.

Critical weaknesses
  • Parameter-efficiency: the authors mentioned parameter-efficiency several times in the paper. However, it is not very clear how the proposed two-branch approaches will be more parameter-efficient than conventional single-branch DiT? More explanation is needed to clarify this.

  • Performance gain from the proposed 2-stage approach: while the presented results are encouraging, the explanation that discrete predictions captured the global structure of action trajectory is not clearly supported by evidence. It is not very clear if the variations in different training runs also contributed to the difference in performance.

Questions for authors
  1. Please clarify the parameter-efficient claim, and compare it with other methods.

  2. Please confirm the details about the fairness of the comparisons, such as whether the same dataset and training protocol were used, and how many fixed random seeds were used. This is important as the success rate improvement does not seem to be statistically significant.

  3. How much of the improvement is gained by the discrete representation, as opposed to the additional one-step reconstruction objective or the change in source distribution?

  4. How is variability across trials handled in the real-world experiments? (also linked to Q1 above) Are the improvements statistically significant?

Reviewer_Yg6JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Few-step action generation requires producing coherent action trajectories under a limited inference budget. This paper introduces Discrete Forcing, which first predicts quantized action tokens and then uses their dequantized values, mixed with Gaussian noise, to initialize a single continuous refinement step. The two branches use partially shared Transformer layers and are trained with discrete reconstruction, flow matching, and an additional endpoint reconstruction loss. The paper evaluates the method on LIBERO, RoboTwin 2.0, RoboCasa-GR1, and four real-robot tasks, with additional studies of inference latency, model scaling, architectural choices, and intermediate action trajectories.

Critical strengths
  • Substantial gains in Simulation Experiment. Discrete Forcing improves RoboTwin 2.0 success from 49.80% to 59.32% across 50 tasks and reaches 56.8% on RoboCasa-GR1, outperforming StarVLA on 21 of its 24 tasks (Tables 2, 5–6). The real-robot evaluation also improves average task progress from 58.5% to 74.9% across precision and dynamic manipulation tasks (Fig. 5). These results provide meaningful evidence beyond a nearly saturated benchmark.

  • A useful performance–latency trade-off. On LIBERO, end-to-end P50 latency decreases from 200.43 ms for 10-step StarVLA to 78.95 ms, while success increases from 96.7% to 97.6% (Table 3). The method also improves over the two-step baseline. Table 4 examines the generation representations, guided source, one-step loss, and branch communication; its predicted-source training control also directly addresses a potential training–inference mismatch.

Critical weaknesses
  • The methodological advance over closely related work is limited. Libra-VLA already uniformly quantizes normalized action coordinates before continuous refinement (Libra-VLA, Secs. 3.2–3.3), so the distinction in Sec. 2 between prior “semantic intents” and two representations of the same action trajectory is overstated. CF-VLA already constructs an action-aware initialization followed by a single refinement step. The identifiable contribution is therefore the particular quantized-source construction and partially shared architecture. Although useful, the current evidence does not establish a sufficiently substantial advance over these existing designs.

  • The specific benefit of discrete guidance is not sufficiently isolated. Table 4 includes a Continuous→Continuous control, but does not specify whether it preserves the source construction, endpoint loss, and branch capacity. These details matter because the one-step loss in Eq. (8) is precisely the flow-matching error at γ=1\gamma=1. The structural analysis in Fig. 4 is also restricted to training demonstrations (Appendix E), and it is unclear whether StarVLA@n denotes a completed n-step prediction or an intermediate noisy state. Consequently, the evidence does not clearly separate a discrete-representation advantage from the benefits of supervised two-stage prediction and endpoint training.

Questions for authors
  1. For the Continuous→Continuous variant in Table 4, how is the coarse prediction trained and converted into the refinement source? Are the endpoint loss, branch capacity, and training budget preserved? For the “w/o Discrete-guided Source” ablation, is the source ϵ\epsilon or (1−α)ϵ(1-\alpha)\epsilon? The supplied implementation’s Gaussian-source option uses ϵ\epsilon, whereas the default guided source is 0.7ϵ+0.3Q−1(Z)0.7\epsilon+0.3Q^{-1}(Z). If this option produced the reported ablation, both the action guidance and Gaussian coefficient change. A control retaining the same Gaussian coefficient during training and inference would better isolate the contribution of action guidance.

  2. In Fig. 4, is StarVLA@n the final output of an n-step solver, or the n-th intermediate state of a longer denoising trajectory? If it is an intermediate state, is it converted into a clean-action estimate before computing the metrics? To complement this offline analysis, what is the closed-loop performance of directly executing the dequantized DF@1 prediction, compared with the full two-stage output from the same checkpoint and under the same replanning protocol? This would directly establish the practical contribution of continuous refinement, particularly on precision-sensitive tasks.

  3. Table 7 specifies 24 layers per execution path, with 2 shared and 22 branch-specific layers, implying 46 distinct Transformer blocks. What are the corresponding baseline depth, width, and parameter counts? Please also clarify whether Table 8 holds total parameter count fixed when changing the number of shared layers.

  4. Could you reconcile the training objectives with the supplied LIBERO configuration? Equation (3) describes masked-token cross-entropy, whereas the implementation also enables a visible-token loss through discrete_visible_loss_weight=0.1. The one-step reconstruction loss is applied to 60% of continuous samples with weight 0.5 and averaged over all continuous samples, a sampling detail not specified in Eq. (9). Please confirm which recipe produced the reported results and ablations, and document the corresponding loss definitions and normalization.

1 comment

Thank you for the clarifications. The parameter counts and source-noise setting resolve my concerns on those points. Two concerns remain:

  1. The supplied implementation uses 56 scalar action tokens in the discrete branch and 8 vector action tokens in the continuous branch. Does the first stage of Continuous→Continuous preserve the scalar-wise tokenization? If not, matching depth and width does not isolate discretization, since token granularity also changes.

  2. The main distinction from related work is now source initialization versus conditioning. Table 4 does not directly compare these two uses of the same discrete prediction. Adding branch communication to the full method is a different control. The response explains the design difference, but its added value over existing coarse-to-fine refinement remains insufficiently established.

The interpretation of Fig. 4(b,c) should also be narrowed: comparing completed discrete predictions with raw intermediate noisy states cannot establish an advantage in clean-action prediction.

The empirical gains are useful, but these remaining concerns about the contribution and its attribution keep me at 2.

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

The paper proposes Discrete Forcing, a two-stage action expert for vision-language-action models. A discrete branch predicts quantized actions from a fully masked sequence; its dequantized output is mixed with Gaussian noise to initialize one-step continuous flow refinement. The branches share early DiT layers and are trained using discrete reconstruction, flow matching, and one-step reconstruction objectives. Evaluation covers LIBERO, RoboTwin, RoboCasa-GR1, three model scales, and four real-robot tasks, together with design ablations.

Critical strengths
  1. The two-stage design addresses a practical quality-latency trade-off. At two function evaluations, LIBERO improves from 95.1% for StarVLA to 97.6% (+2.5 percentage points), RoboTwin from 49.80% to 59.32% (+9.52 points), and RoboCasa-GR1 from 43.9% to 56.8% (+12.9 points). Table 3 reports P50 action-only and end-to-end latency, with approximately 5.84x and 2.54x speedups over the ten-step baseline, respectively. The end-to-end gain over the two-step comparator is more modest (93.29 versus 78.95 ms, approximately 1.18x), but measured system-level speed alongside task gains is the strongest reason for acceptance.

  2. The evidence extends beyond one benchmark. Table 4 includes generation-order, source-guidance, one-step-loss, and predicted-source-training ablations; further experiments vary the noise mixture, shared layers, and three model scales. The predicted-source training result (95.3 versus 95.2 on LIBERO-Long) partly addresses the training/inference source shift, although the difference is small. On four real-robot tasks with a pi0 backbone, mean Task Progress is 74.9 versus 58.5 (+16.4 percentage points), adding useful transfer evidence while measuring progress rather than binary task success.

Critical weaknesses
  1. The method and its causal attribution need tighter specification. Equations (6)-(7)/(13) write the continuous field as f(A_gamma,Z,o), but it is unclear whether Z is a separate conditioning input or acts only through the initialized source. Training normally constructs this source from ground-truth quantized actions, whereas inference uses predicted tokens; the predicted-source ablation is reassuring but lacks uncertainty. Table 4's Continuous-to-Continuous control scores 91.8 versus 95.2 for Discrete Forcing, yet its source construction, intermediate supervision, architecture, and training budget are insufficiently specified to isolate discreteness from the wider two-stage design. The claimed parameter match also remains unauditable: Table 7 specifies 24 layers per path, two shared and 22 branch-specific, implying 46 distinct transformer blocks; baseline depth, width, and component counts are missing. If per-path depth is fixed, changing the number of shared layers in Table 8 would also change the number of distinct blocks unless another capacity control is applied. None of this proves an unfair comparison, but it limits attribution of the gains to discrete guidance.

  2. The theoretical and experimental boundaries of the mechanism claim are not fully established. The flow-matching equations are internally plausible, but no argument shows that token prediction necessarily supplies global action structure, that one continuous step is generally adequate, or that the observed three-scale trend will persist beyond the tested range. Figure 4's structural analysis uses 200 action chunks from training demonstrations with bootstrap resampling; it supports an in-sample diagnostic, not a held-out structural claim, and chunk-level resampling may miss within-task dependence. The main success-rate comparisons and small ablations do not report enough training-seed or episode-level uncertainty to judge stability. The real-robot study has 20 trials per task and reports Task Progress, but trial-level outcomes and matched training/control details are needed to interpret the +16.4-point mean progress difference. These are limits on confidence and mechanism, not a negation of the larger observed benchmark gains.

Questions for authors
  1. Please give the action-expert and total component-level parameter counts, baseline depth/width, and training compute for each principal comparator. In Table 8, which of distinct parameters, per-path depth, and training compute are fixed as sharing varies? If capacity changes, can the performance claim be separated from that effect?

  2. In Equations (6)-(7)/(13), does Z enter f(A_gamma,Z,o) as a separate conditioning input, or only through continuous-source initialization? Please specify the exact training and inference pathways, branch-assignment probabilities, and paired-Beta time-sampling transformation.

  3. For Table 4's Continuous-to-Continuous control, which source construction, intermediate targets, one-step supervision, architecture, and training budget match Discrete Forcing? Can the predicted-source ablation and Figure 4's structure diagnostic be reported across seeds and held-out trajectories or tasks, with task-aware uncertainty where applicable?

  4. What training-seed and evaluation-episode protocols underlie the main comparisons and small ablations, and what uncertainty estimates are available? For the real-robot comparison, please clarify matched data, training and control-loop settings, and show per-trial or per-task Task Progress alongside full-task success if measured.

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

Related papers

Discussion (0)

Sign in to comment.