acceptodds
Under review as a conference paper at ICLR 2027

Training-Free Adaptive Time Grids for Flow-Based Vision-Language-Action Models

Abstract

Flow-based vision-language-action (VLA) policies generate action sequences by numerically integrating a learned probability-flow ordinary differential equation. Under a limited inference budget, the choice of time discretization can substantially affect control quality, especially after reinforcement-learning (RL) post-training, which may change the velocity field away from the geometry induced by conditional flow matching. We study this problem through a training-free, inference-time perspective. We first separate three quantities that are often conflated: marginal entropy rate, local velocity-field variation, and numerical integration error. Based on this analysis, we propose Entropy- and Geometry-aware Time Grids (EATG), which estimates trajectory-dependent complexity from a calibration rollout and constructs a constrained non-uniform time grid under an explicit inference-cost budget. The method combines entropy-rate proxies with along-trajectory velocity variation, supports fallback to uniform discretization when the signals are unreliable, and requires no parameter updates or auxiliary networks. We evaluate EATG on flow-based VLA policies before and after RL post-training across simulated and real-world manipulation tasks. Our experiments compare entropy-only, geometry-only, fixed-grid, and equal-budget numerical baselines, and analyze robustness across initial noise realizations. The results show when adaptive discretization improves closed-loop success and when reducing ODE integration error alone is insufficient, highlighting the distinction between generative numerical accuracy and task-level control quality.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, flow policy

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7FQAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes EATG, a training-free inference method that replaces uniform time discretization in flow-based VLA models with an adaptive grid. The method runs a calibration rollout on a uniform grid to estimate a per-timestep complexity score from two signals, entropy evolution and velocity-field geometry, then reallocates the inference budget toward high-complexity regions in a second rollout. The motivation is that flow trajectories have non-uniform complexity, and that RL post-training distorts the transport geometry so entropy alone is insufficient. EATG is applied to SFT and RL-finetuned policies on LIBERO, ManiSkill, CALVIN, Meta-World, and a Franka robot, reporting gains over uniform sampling, with the largest on real-world tasks.

Critical strengths
  • The problem is real. Flow-based VLA inference is expensive, and uniform grids are a crude default that ignores where the trajectory is actually hard.
  • The observation that RL post-training distorts the conditional-flow transport geometry, so an entropy-only schedule derived from CFM assumptions can mislead, is useful and testable.
  • The method is genuinely plug-and-play: no training, no weight modification, applied at inference time.
Critical weaknesses
  1. The comparison is not matched-cost, and the paper's own ablation suggests the gain comes from extra computation rather than smarter allocation. EATG runs two rollouts (calibration plus inference), so it spends roughly twice the evaluations of a uniform grid, yet every comparison is uniform-at-N versus EATG-at-more-than-N. No experiment gives the uniform baseline the same total budget. Table 5 already shows the risk: a geometry probe using four times the evaluations beats uniform by only 0.15 points (0.16 claimed in the paper). The paper itself notes that these results "motivate matched-cost component ablations before attributing any gain specifically to the combination." That matched-cost comparison is missing, and it is the comparison the claim depends on.

  2. The entropy and geometry components are not isolated, and the theory does not identify the mechanism. There is no ablation of entropy-only versus geometry-only versus the combination, so it is unclear what the method adds. The analysis is a rearrangement of mutual information plus an identity that holds only under the linear-interpolation assumption of the conditional flow, not the learned marginal flow the model actually follows. The estimator used in calibration is ambiguous (the closed form of Eq. 13 versus the Hutchinson divergence of Eq. 15), the normalization is unspecified, and the key hyperparameter λ in the α-combination is given no value and is not ablated.

  3. The real-world result is the least controlled part of the evaluation. Four tasks, 40 trials each, one robot, no seeds or error bars, and the baseline's step count is not reported. A timestep scheduler moving Cap Alignment from 30% to 75% is hard to explain as better numerical integration; it more likely reflects a baseline that is under-discretized, in which case the two-pass EATG is simply spending more compute. This is the same cost problem as Weakness 1, applied to the strongest reported numbers.

  4. No existing adaptive scheduler is used as a baseline, and the novelty is correspondingly limited. The paper cites entropy schedulers (Stancevic et al., Trentini et al.) and geometry-aware VLA inference (ProbeFlow, AdaVLA) but compares only against uniform sampling. Because the contribution is the entropy-plus-geometry combination for RL-finetuned VLAs, the method needs to beat these existing schedulers at matched cost, not just uniform. Without that, the contribution amounts to combining two known techniques, which the paper's own related-work discussion concedes.

  5. A minor presentation point. Figures should be redrawn as vector graphics, and table captions should be placed above the tables rather than below them.

Questions for authors
  1. What happens when the uniform baseline is given the same total budget as EATG, including the calibration rollout? Does the adaptive grid still improve the success rate?
  2. What does each component contribute in isolation: entropy alone, geometry alone, and the α-combination? What is the value of λ, and which entropy estimator (Eq. 13 or Eq. 15) is actually used in calibration?
  3. For the real-world experiments, what is the baseline's step count and schedule, and do the large gains survive across seeds and more trials? How does a timestep scheduler explain swings of this size?
  4. How does EATG compare with existing adaptive schedulers at matched cost, and what is the specific novelty beyond their combination?
Reviewer_37jROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Entropy- and Geometry-aware Adaptive Time Grids (EATG), a training-free inference-time scheduling method for flow-based VLA models. The motivation is that uniform temporal discretization relies on the assumption that the uncertainty and geometric complexity are constant along the transport trajectory, while in fact they are not. The key framework of EATG is that it first performs calibration rollouts to estimate trajectory complexity based on entropy (which measures the uncertainty) and local velocity-field geometry, and then constructs the time grid accordingly for the final flow integration. The authors evaluate the method on both SFT and RL-finetuned VLA policies, both in-domain and out-of-domain, and several real-world manipulation tasks. The reported results show consistent improvements, especially on OOD ManiSkill tasks and real-world experiments.

Critical strengths
  1. The empirical results are strong and cover a relatively broad range of settings. The improvements on ManiSkill OOD tasks and the real-world experiments are especially strong, providing a good motivation for non-uniform scheduling, especially for RL-finetuned VLA policies. 2.The paper has an intuitive and interesting insight about how RL post-training can change the geometry of velocity field, and how it affects transport trajectory.
Critical weaknesses
  1. A central claim of the paper is that entropy alone is insufficient because RL changes the velocity field geometry, but there is no clear ablation comparing entropy-only scheduling, geometry-only scheduling, and the full method.
  2. The experimental setting is not described clearly enough to interpret or reproduce several of the main results. Specifically, Tables 1, 2, and 6 do not report the number of NFEs used by the baseline and EATG, including both the NFEs for calibration and for the final integration. So it is unclear whether the reported comparisons use the same final integration budget. Several additional implementation details, such as the choice of calibration-grid size and some scheduling hyperparameters, are also not clearly specified either, making the experimental setup hard to reproduce.
Questions for authors
  1. What exactly is included in the wall-clock measurement reported in Table 4? EATG increases the number of velocity-field evaluations from 5 to 10, but the reported runtime only increases from 28s to 30s. The paper explains that the calibration phase only requires forward evaluations without gradient computation, but standard inference also usually does not require gradient computation. Also, Table 4 reports 5 NFE for uniform sampling while section 4.2.3 reports 10 NFEs for uniform sampling.
  2. Clarification on the relationship between entropy evolution, complexity score, and integration steps: Section 3.4 claims that slow entropy contraction corresponds to higher uncertainty and therefore requires finer temporal discretization. However, according to Eq. (15), a larger local divergence magnitude means a larger complexity score, while a larger complexity score means smaller temporal spacing. These descriptions seem to suggest opposite relationships between the entropy change and integration step size.
  3. See critical weaknesses.
Reviewer_6fcsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes EATG, a training-free way to pick a non-uniform time grid for flow-based VLA policies. For each inference, it first runs a calibration pass on a uniform grid with the same initial noise, records a complexity score per interval (a mix of a divergence-based "entropy rate" proxy and the temporal change of the velocity field, weighted by trajectory curvature), and then re-integrates on a grid that puts more steps where the score is high. The motivation is that RL post-training bends the velocity field, so entropy alone isn't enough. Results are reported on LIBERO (pi0.5 SFT), ManiSkill/Meta-World/CALVIN with RLinf Flow-SDE checkpoints (IND and OOD), and four real-world Franka tasks with pi0.

Critical strengths

The question is reasonable: whether RL post-training changes the flow geometry enough that standard (or entropy-based) step schedules stop being a good fit is a fair and practical thing to look at, and a plug-in inference-time fix would be useful if it works. The real-robot numbers (50% -> 85%) would be striking if they hold up under a proper comparison.

Critical weaknesses

There's no matched-compute comparison, and the paper's own ablation undercuts the main claim. EATG runs a calibration pass and then a second pass, so it uses roughly twice the NFEs of the uniform baseline it's compared against (and Table 4 vs. its text even disagree on the uniform NFE). The obvious baseline, uniform with the same total NFE, is missing, and so are entropy-only, geometry-only, the entropic time scheduler it builds on, ProbeFlow, and AdaVLA. Meanwhile Table 5 shows the geometry probe at 4x the NFE is basically tied with uniform (87.0 vs. 86.9), and the text itself says matched-cost ablations are needed before attributing gains to the method. As it stands, I can't tell whether any of the improvements come from the grid rather than just more compute. The gains themselves are also mostly small (e.g. +0.8 on LIBERO) with no seeds or error bars, while the real-robot jump is large and comes with no protocol details.

The paper doesn't feel finished. The theory is mostly standard identities (Theorem 1 is the entropy chain rule, and H(X_t|P) = 0 doesn't really make sense for differential entropy), and the "entropy" term actually used is the absolute divergence at a single sampled point, not an entropy rate. The motivating Fig. 2 shows only five PCA-projected trajectories per model, and the RL ones are still mostly straight (one bends slightly near the end) and mainly differ in step spacing, which is weak support for the "highly curved" claim; the radar chart's axes (e.g. Plate, EEPose) also don't match Table 3. On top of that there's an unfinished sentence at the end of Sec. 3.4, a broken pi0.5 reference, typos in headings and table rows, and the OpenReview abstract describes ablations (entropy-only, geometry-only, equal-budget, noise robustness) that I couldn't find in the PDF.

Questions for authors
  1. What do you get with a uniform grid at the same total NFE as EATG (calibration + second pass), on the same benchmarks? This is the key comparison for me.
  2. Can you report entropy-only and geometry-only variants, plus the entropic time scheduler, at matched cost?
  3. For the real-robot results: how many steps does the baseline use, were trials interleaved and started from the same initial conditions, and how was success judged?
  4. The OpenReview abstract mentions ablations and noise-robustness analyses that aren't in the PDF. Is there a missing version or appendix?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.