- 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
-
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.
-
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.
-
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.
-
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.
-
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
- 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?
- 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?
- 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?
- How does EATG compare with existing adaptive schedulers at matched cost, and what is the specific novelty beyond their combination?