- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes AC-HFlow, a flow-matching policy for visuomotor action-chunk prediction that comprises two components: 1) Action-Consistent Attention (ACA) projects timestep features into a learned joint–time grid and applies a union mask combining causal temporal attention within each slot and directed attention between slots at each timestep. 2) History Flow Matching constructs a flow source using historical action means and standard deviations, followed by affine standardization.A positional-encoding variant, GRoPE, applies successive rotary transformations along the temporal and joint-slot indices.
Critical strengths
-
A simple and well-motivated architectural idea. Explicitly exposing the joint–time structure of an action chunk to the velocity network is a reasonable question for chunked generative policies. ACA is a lightweight, zero-gated, and drop-in layer that does not change the training objective, so it would be easy to adopt in existing flow and diffusion policies.
-
Transparent reporting of the experimental setup. The appendix provides per-task results for every step count, the LIBERO-Plus scene IDs and counts, and an explicit discussion of what the diagnostics do not claim. This makes it possible to check the claims in detail, and the authors are honest that the cross-family comparisons are not compute-matched.
Critical weaknesses
Weakness 1: The History Flow Matching source does not carry the history information it is credited with.
After the standardization in Eq. 11, the source simplifies to . If and are sampled per element, as the text suggests, then exactly, and the source is identical to standard flow matching. If is instead shared across the horizon, the history mean still cancels out, and the only remaining effect is a per-control temporal correlation. In either case, the claims that the source is "aligned with the previously executed motion" and that it "shortens the transport distance" are not supported by the formulation. There is also a mismatch between training and inference: the buffer stores ground-truth history during training but the policy's own predicted chunks at test time.
The paper's own results are consistent with this concern. In Table 9, the union model reaches 78.1% at a single step both with and without history, and adding history to ACA improves the average by only 0.9 points (78.6 → 79.5). This weakens the second contribution, which credits the history-shaped source with enabling few-step inference.
Weakness 2: Unconvincing empirical evidence for the claimed advantages.
Much of the reported gain is measured against the FM-DiT baseline, which scores 0% on three of the seven tasks (24.3% average), while DDPM-DiT, from the same backbone family, reaches 73.8%. Every single modification (history only, joint-only ACA, time-only ACA, or union ACA) restores performance to roughly 78%. Given Weakness 1, this pattern points to an issue with the baseline configuration rather than to the proposed inductive bias. Likewise, the joint-only, time-only, and union variants perform almost identically, so the ablation can not show that the specific partial order matters.
For example, the paper lacks controlled comparisons between:
- union-masked ACA and dense (unmasked) attention on the same grid;
- the history-shaped source and plain Gaussian noise under an otherwise identical code path;
- AC-HFlow and the baselines at the same small step budgets (e.g., DDIM at 10 steps rather than 90).
Questions for authors
-
The second contribution centers on the history-shaped source. Could the authors clarify how is sampled in Eq. 10 and state the resulting distribution of ? Would it be possible to report a variant that uses as the source with an otherwise identical code path, to verify that the gain actually comes from the history statistics?
-
The FM-DiT baseline fails completely on four tasks. Could the authors clarify whether it is exactly AC-HFlow with ACA and history removed (same normalizer, AdaLN conditioning, EMA, solver, and parameter count), and explain why it performs so much worse than DDPM-DiT? Reporting parameter and FLOP counts for all policies would also help.
-
The paper does not compare the union mask against dense attention on the joint–time grid or against a parameter-matched extra attention layer. Would it be possible to report such comparisons, so that the effect of the proposed relational structure can be separated from the added capacity?
-
The joint factors are learned linear projections of the hidden feature, and on LIBERO they factorize a 7-D OSC end-effector action rather than physical joints. Could the authors clarify why the slot ordering is meaningful, and whether performance changes when the order is randomly permuted or when is set to the action dimension?
I would be willing to raise my score if the authors can adequately address my concerns.