acceptodds
Under review as a conference paper at ICLR 2027

PACT-VLA: Speculative Action-Chunk Verification with Pre-Commit Observations

30 views

Abstract

Vision-language-action (VLA) policies are commonly deployed with asynchronous inference in real-world settings, where the next action chunk is generated while the robot is still executing the previous one. By the time the new chunk is committed, the robot and scene have changed, so the chunk is conditioned on a stale observation and task success drops sharply as this inference delay grows. Generating multiple candidate chunks could let newer observations inform the final choice, but existing multi-candidate methods either rank candidates with the stale inference-start observation or repeatedly switch between candidates during execution, which can make the executed action sequence discontinuous. Our key insight is that inference delay can be used as decision time. While candidate generation is still underway, candidate selection can reuse the VLA itself to choose the complete action chunk that best fits a later observation. We propose PACT-VLA, a training-free framework that generates candidates from the inference-start observation and selects one using this pre-commit observation. Its scoring function, FlowRes, measures the residual between each candidate's implied flow direction and the velocity predicted by the frozen policy under the pre-commit observation, requiring no separate verifier. A pipelined implementation overlaps observation processing with candidate generation and batches scoring, so selection adds no control-step delay on the evaluated hardware. At a four-step delay, PACT-VLA improves average success by 15.0 percentage points over RTC on Kinetix and by 5.2 points over single-candidate asynchronous execution with GR00T N1.7 on LIBERO, and it further improves existing delay-robust methods. On three real-robot tasks, it raises average success from 60% to 82% over naive asynchronous execution with the same frozen π₀.₅ checkpoint.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Action Chunking, Inference Delay, Test-Time Action Selection, Flow Matching, Speculative Verification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fWooOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the stale-observation problem in asynchronous execution of action-chunking flow-based VLAs. PACT-VLA generates N candidate chunks from the inference-start observation oto_t, and selects one using the pre-commit observation ot+d−1o_{t+d-1}, i.e., one control step before the planned chunk switch. Selection uses FlowRes, the flow-matching velocity residual of each candidate under the frozen policy conditioned on ot+d−1o_{t+d-1}, computed with shared verification noise over a few flow times. Encoding of ot+d−1o_{t+d-1} overlaps with candidate generation and all candidates are scored in one batch, so on the evaluated hardware the cycle stays within the same 30 Hz delay bucket (d=2) as single-candidate generation. Experiments on Kinetix (vs. many delay-handling baselines, and on top of VLASH/DEFLECT/REMAC), LIBERO (π0.5 and GR00T N1.7), and three real-robot bimanual tasks with π0.5 show gains that grow with delay (e.g., +15.0 pts over RTC at d=4 on Kinetix; 60%→82% on real robots).

Critical strengths
  • Simple, training-free, and complementary. The method reuses the frozen policy as the verifier, requires no extra model, and improves several existing delay-robust methods (Table 2). This makes it practically attractive and easy to adopt.
  • Broad empirical coverage with a targeted control. Table 1 compares against a wide set of delay-handling methods under a common checkpoint and protocol. The same-pool comparison that changes only the scoring observation (oto_t vs. ot+d−1o_{t+d-1}; Tables 7–9, Old-FlowRes on the robot) directly supports the paper’s central claim, and the gain consistently grows with delay.
Critical weaknesses
  1. Limited explanation of why the method works. The paper describes how PACT-VLA is implemented in detail but gives little analysis of why FlowRes under the pre-commit observation should identify better chunks. The authors themselves note that FlowRes "is a policy-consistency score rather than a direct estimate of task success" and defer its justification to experiments (l.272–274). There is no discussion of its relation to conditional likelihood, no analysis of when the candidate pool (generated from oto_t) contains a chunk that fits ot+d−1o_{t+d-1} (e.g., ablations on NN or candidate diversity), and no qualitative examples of selected vs. rejected candidates. Relatedly, the single-decision gain in the fixed-pool evaluation is small (+0.56 pts, D.2), while episode-level gains reach +15 pts (Table 9). Without an explanation of this gap, it is hard to attribute the improvement to the stated mechanism.

  2. Unclear temporal alignment in FlowRes. Candidates are generated from oto_t, so index jj corresponds to step t+jt+j. The velocity predicted under ot+d−1o_{t+d-1} presumably corresponds to a chunk starting at t+d−1t+d-1. The paper does not state whether candidates are re-indexed before computing Eq. 6. This matters most on Kinetix, where the largest gains are reported: with RTC-guided proposals the first dd actions are pulled toward the committed prefix and are nearly identical across candidates, yet with wj=0.5jw_j = 0.5^j, H=8H=8 and d=4d=4 they carry ~94% of the score weight. If no re-indexing is done, it is unclear what the score actually discriminates. This is a soundness question rather than a presentation issue.

Questions for authors

Clarifications (could change my assessment)

  1. Is the candidate re-indexed (e.g., comparing Aj+d−1A_{j+d-1} with the prediction at index jj) before computing FlowRes? Please specify the delay-simulation protocol for each benchmark (a time-indexed queue as in RTC, or executing each new chunk from index 0). For RTC-guided candidates, is the verification query also conditioned on the committed prefix?

  2. How do you reconcile the small single-decision gain in D.2 (+0.56 pts) with the large episode-level gain in Table 9 (+15 pts at d=4)?

  3. Statistical reliability of the real-robot results. With 50 trials per method, the standard error of each success rate is roughly 5–7 pts. Over Old-FlowRes, the averaged gain (+17.3 pts, z≈3.5) and Pouring Coffee Beans (+28 pts, z≈3.4) are clearly significant, but Toasting Bread (+14) and Loading Dishwasher (+10) are within about 1.5 standard errors of the difference. The varying baseline rankings across tasks are also likely within noise; for example, Old-FlowRes is best among baselines on Dishwasher but below Naive Async on Coffee Beans. Please report:

    • confidence intervals;
    • the trial protocol (ordering, initial-state matching);
    • how the VLASH checkpoint was trained;
    • how PACT's hyperparameters (NN, the 25-step scoring window, flow times) were chosen, and whether they were tuned on the evaluation tasks.
  4. What is the effective delay dd on the robot? With d≈2d\approx2, the pre-commit observation is only ~33 ms newer than oto_t, yet the gain over Old-FlowRes (+17.3 pts) is much larger than the d=2 gains in simulation (+1.1 on LIBERO, +4.0 on Kinetix).

Missing comparisons (would strengthen the paper)

  1. LIBERO and the real-robot experiments compare only against single-candidate async (plus VLASH on the robot). Comparisons against RTC in these settings, and against test-time selection methods (e.g., RoboMonkey, MG-Select, τ0-WM), would help. Prior work has also used the latest observation during chunk execution through lightweight learned residual correction without candidate switching (e.g., arXiv:2509.23224). How does pre-commit selection compare with such per-step correction, in both success and latency?

  2. Sensitivity to NN, MM, the flow-time set, and wjw_j, which differ across models (Table 6). Since simulations fix dd externally and do not reflect the added compute, please also report latency for N=8N=8 and for GR00T.

1 comment

We thank the reviewer for recognizing the training-free design and controlled comparisons. We have highlighted the revisions in blue and hope our responses address your concerns.

Comment/Question:

Limited explanation …

Answer: FlowRes favors actions consistent with demonstrated behavior under the updated state, reducing the mismatch caused by stale observations. Following this helpful suggestion, we expanded Section 3.2 to discuss the mechanism and conditional likelihood. The flow-matching objective provides a likelihood-related interpretation, while FlowRes uses its residual as a policy-consistency surrogate rather than an exact likelihood estimate. Appendix D.2 (Table 10) shows that 14.1% of decisions with large candidate-outcome differences contribute 56.8% of the gain. For such candidate pairs, pre-commit scoring identifies the better candidate 56.2% of the time, compared with 51.7% using the original observation. Appendix G links larger gains to higher single-candidate failure rates, while Appendix J (Fig. 9) shows candidate rankings and outcomes.

Comment/Question:

Unclear temporal alignment …

Answer: Kinetix candidates are re-indexed before FlowRes scoring, so the 94% positional weight falls on post-switch actions rather than the committed prefix. We remove the first d already-committed positions, which the new chunk cannot change, then shift the executable suffix forward and zero-pad the sequence. Padded positions retain the remaining weight and can contribute nonzero, candidate-dependent residuals. We clarify this in Appendix B.

Comment/Question:

Is the candidate re-indexed …

Answer: Verification uses the frozen policy without RTC prefix guidance, while action indexing follows each benchmark's execution protocol. Kinetix uses RTC's time-indexed queue. LIBERO executes new chunks from index 0, while the real robot uses LeRobot's timestep-based queue merging. We use o(t+d−1) as the latest observation that leaves one control step for verification before the switch at t+d. The execution protocols are detailed in Appendices B and H.

Comment/Question:

How do you reconcile …

Answer: Full-rollout gains reflect repeated selections, while the fixed-pool experiment isolates one decision. Appendix D.2 (Table 11) varies selection frequency: selecting only once adds 0.7 points, every second decision adds 8.0, and every decision adds 15.6. This demonstrates accumulation across decisions.

Comment/Question:

Statistical reliability …

Answer: The 17.3-point average gain over Old-FlowRes is statistically supported, with a 95% CI of [+7.7, +27.0] (Appendix H, Table 13). Task-level CIs are also reported. All methods use LeRobot's default asynchronous execution settings, with rotated trial blocks and matched initial placements. Appendices E and H detail checkpoint training and deployment parameters, which were not tuned on the real-robot tasks.

Comment/Question:

What is the effective delay …

Answer: The larger real-robot gain over simulation is explained by the longer execution horizon affected by each selection, despite both operating at d=2. At 30 Hz, the observation is 33 ms newer. Robot chunks are requested every 25 steps, and each choice shapes ~1.6 s of motion, versus replanning every 2–8 actions in simulation (Appendix H).

Comment/Question:

Comparisons against RTC …

Answer: We have added RTC comparisons on LIBERO and real robots, alternative-selector controls, and A2C2 success and latency measurements. On LIBERO, RTC + PACT-VLA improves success from 93.2% to 94.2% at d=4 and from 73.8% to 76.1% at d=20 (Appendix M). On real robots, PACT-VLA reaches 82.0% versus RTC's 67.3% (Fig. 5). On GR00T, PACT-VLA achieves 56.4% versus 51.5% for τ0-WM's RCS (Table 7). RoboMonkey reports ~650 ms for 16 candidates under a different setup, while MG-Select's action-token score does not directly apply to flow policies (Appendix D). For A2C2, reproduced using the authors' code on our shared Kinetix policy, PACT-VLA achieves higher success at all four delays, from 91.1% versus 78.8% at d=1 to 77.2% versus 75.8% at d=4 (Table 19). On RTX 5090, P99 latency is 43.4 ms for generation plus the first A2C2 correction versus 60.2 ms for PACT-VLA with N=4, both retaining d=2 at 30 Hz (Table 20). A2C2's adapted correction head runs at every control step, while PACT-VLA verifies once per action chunk and requires no additional training.

Comment/Question:

Sensitivity to …

Answer: The revision reports sensitivity, candidate-count, and GR00T P99 latency results. Fig. 7 varies flow times, step weights, scored positions, and action dimensions. Every tested setting beats its no-selection baseline. Appendices G/I vary N. Tables 4/16 cover N=1,2,4,8 on three GPUs. On RTX 5090, GR00T takes 34.1 ms (N=1) versus 62.5 ms (N=4), both d=2 at 30 Hz. Appendix K covers 25/20 Hz.

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

This paper addresses the stale observation problem. It proposes PACT-VLA, which selects a candidate by also considering later observations with a four-step delay. The method achieved better performance than asynchronous execution with GR00T N1.7 on LIBERO.

Critical strengths
  1. The work noticed the problem of stalled observation and addressed it with adding later observations, which is missed by a lot of other methods. The work also obtained no additional control-step delay on the tested hardware on RTX 5090.
  2. It requires no additional trainning needed for the trajectory generation and verification. And the performance is evaluated across diverse benchmarks with positive result.
Critical weaknesses
  1. The paper needs to establish why a lower FlowRes score corresponds to better performance. It measures the consistency between the frozen policy's intended trajectory on the new image and the candidate's proposed trajectory. Such consistency doesn't guarantee a better success rate, and I need a deeper analysis of this choice.
  2. The candidate trajectories are still outdated. If the scene changes a lot, then the candidate trajectory cannot fix the issue. The newer information cannot be used to revise the candidate.
  3. The method still introduces additional latency beyond policy prediction. Even though the author tested that the method doesn't add additional control-step delay on RTX 5090, this delay depends on the hardware.
Questions for authors
  1. Why did you choose FlowRes as an indicator for choosing the candidate?
  2. The method would work if these candidates are different from each other, and the policy can choose one among the candidates. How is candidate diversity in the work, and what is the tradeoff between candidate diversity, candidate number, and success rate?
4 comments

We thank the reviewer for the constructive feedback and for recognizing the value of using later observations to address stale-action execution, together with the training-free nature and broad evaluation of PACT-VLA. We have highlighted the revisions in blue and hope our responses address your concerns.

Comment/Question:

The paper needs to establish why a lower FlowRes score corresponds to better performance …

Answer: Lower FlowRes scores indicate greater consistency with the frozen policy's learned velocity field under the updated observation. Since the policy is trained on task demonstrations, this consistency provides a useful signal for selecting candidates that better fit the updated state. We have added this explanation to Section 3.2. We verify this connection against actual execution outcomes. In the fixed-pool analysis (Appendix D.2, Table 10), for candidate pairs with large outcome differences, pre-commit scoring identifies the better candidate 56.2% of the time, compared with 51.7% under the generation observation. On the real robot, changing only the scoring observation improves success from 64.7% to 82.0%.

Comment/Question:

The candidate trajectories are still outdated …

Answer: A scene change does not necessarily invalidate all sampled candidates. When viable alternatives remain, the newer observation can guide the selection of a suitable action without another generation cycle. Our fixed-pool analysis (Appendix D.2, Table 10) shows that candidate outcomes differ in 80.4% of decisions, with the 14.1% showing the largest differences accounting for 56.8% of the selection gain. Crucially, changing only the verification observation improves Kinetix success from 62.02% to 77.16% at d=4 (Table 9), demonstrating the value of selecting from the existing pool. When a change requires entirely new actions, regeneration is necessary. PACT-VLA complements rather than replaces such approaches, as supported by its consistent improvements over delay-robust proposal methods in Table 2.

Comment/Question:

The method still introduces additional latency …

Answer: PACT-VLA accommodates candidate verification within the existing asynchronous inference window, preserving the same effective control-step delay as single-candidate generation at N=4 and 30 Hz across all three evaluated GPUs. This is achieved by overlapping pre-commit observation encoding with candidate generation and scoring all candidates in a single batch. On RTX 5090, P99 latency increases from 35.4 ms (N=1) to 60.2 ms (N=4), but both remain at d=2 (Table 4). Appendix L (Table 16) confirms the same delay buckets for GR00T N1.7 across all three GPUs. The candidate count provides a direct latency–performance trade-off. At N=8, the delay increases to d=3 on two GPUs, while Appendix K reports the corresponding effects at 25 and 20 Hz. We therefore use N=4 for real-robot deployment, balancing selection performance with the available control-step budget.

Comment/Question:

Why did you choose FlowRes ...?

Answer: We choose FlowRes because the flow-matching training objective provides a natural verification signal that can be evaluated under updated observations using the frozen policy, without a separately trained verifier. FlowRes measures the candidate's velocity residual under the newer observation (Eqs. 5–6). Appendix D (Table 7) compares alternative selection rules under the same N=8 candidate budget on GR00T at d=4. PACT-VLA achieves 56.4% success, compared with 52.2% for random selection, 52.6% for generation-observation FlowRes, and 51.5% for re-denoising consistency (RCS).

Comment/Question:

How is candidate diversity in the work, and what is the tradeoff ...?

Answer: Candidate diversity provides alternative actions for selection. Increasing the candidate count improves success by expanding these alternatives, but the gains diminish while inference cost increases. Diversity arises naturally from independent noise samples of the frozen flow policy, without explicit diversity optimization. Appendix G (Fig. 8) measures it by the standard deviation of candidate actions. At d=4 on Kinetix, increasing N from 2 to 4 to 8 raises candidate spread from 0.20 to 0.29 to 0.33, while success increases from 71.3% to 77.2% to 81.0% (61.5% at N=1). Importantly, selection benefits from meaningful differences in candidate outcomes, not merely larger action-space distances. Candidate spread does not positively correlate with success gains across levels (ρ=−0.09), while Appendix D.2 (Table 10) shows that decisions with substantial candidate-outcome differences contribute 56.8% of the selection gain. We use N=4 for deployment because it retains nearly all the N=8 gain on π0.5 LIBERO (+3.2 vs. +3.5 points at d=4, Appendix I), while avoiding the additional control-step delay that N=8 incurs on two evaluated GPUs (Table 4).

Thank you for the detailed responses and additional experiments. They clarify several concerns, but I still have the following reservations.

  1. There isn't enough evidence that a lower FlowRes score can identify better candidates. Even for candidates with large outcome differences, the ranking accuracy is only 56.2%, which is modest. And the average branch success only increased from 0.55% over the old observation.
  2. The candidates are still generated from outdated observations. I agree with the author that newer observations can help select a suitable candidate when one exists. Yet if the scene changes, the method cannot revise the candidates.
  3. Additional computation still increases control-step delay. The measurements across three GPUs address part of my concern. Yet the reported results in the supplementary section K, Table 15, show that verification adds a control step and can outweigh the selection gains. Therefore, the practical benefit depends on having enough time available for verification.

Other asynchronous dual system designs, such as AsyncVLA, respond beyond a fixed candidate pool by receiving updated observations and generating actions based on them. Such systems also do not have to wait for the large VLA and action policy runs in real time. Although PACT-VLA avoids additional training, its ability to use newer observations remains limited to selecting existing candidates.

Hence, I maintain my original score.

We thank the reviewer for the follow-up and for acknowledging that our additional experiments clarify several concerns. We would like to further clarify three points.

Comment/Question:

Even for candidates with large outcome differences, the ranking accuracy is only 56.2% ... the average branch success only increased ...

Answer: The benefit of FlowRes comes from repeatedly improving candidate selection throughout execution, rather than relying on a large gain from any single decision. For candidate pairs with large outcome differences, pre-commit scoring identifies the better candidate 56.2% of the time, compared with 51.7% using the original observation (Table 10). The fixed-pool analysis reports a +0.56 percentage-point gain for one selection (Appendix D.2). Table 11 directly validates how these gains accumulate. Selecting only at the first decision improves episode success by 0.7 points, while selecting every eighth, fourth, and second decision yields gains of 2.1, 4.0, and 8.0 points, respectively. Selecting at every decision raises success from 61.4% to 77.0% (+15.6 points). On real robots, changing only the scoring observation also improves success from 64.7% to 82.0%. These results establish the value of repeated pre-commit selection at the episode level.

Comment/Question:

The candidates are still generated from outdated observations ... the method cannot revise the candidates ... Other asynchronous dual system designs, such as AsyncVLA ...

Answer: PACT-VLA addresses stale-action mismatch through updated-observation selection without retraining the policy, while generating entirely new actions is a complementary capability. We agree that regeneration is necessary when no viable candidate remains. However, selecting among existing candidates can still improve execution, as demonstrated by our experiments. In particular, Table 2 shows that PACT-VLA further improves three delay-robust proposal methods, VLASH, DEFLECT, and REMAC, by 2.8, 1.5, and 1.5 points on average. AsyncVLA uses an additionally trained reactive module to generate actions from updated observations, whereas PACT-VLA reuses the original frozen policy without an additional model. These approaches therefore address stale observations through different and complementary mechanisms.

Comment/Question:

Additional computation still increases control-step delay ... Table 15 shows that verification adds a control step and can outweigh the selection gains.

Answer: Additional computation does not necessarily increase control-step delay when verification fits within the existing inference budget. At 30 Hz, PACT-VLA with N=4 maintains d=2 across all three evaluated GPUs, matching single-candidate generation (Table 4). Importantly, our real-robot experiments use this same 30 Hz, N=4 configuration and achieve 82.0% success with verification included in the deployment pipeline (Section 4.2, Appendix H). We agree that the benefit is hardware- and control-frequency-dependent. Appendix K explicitly reports configurations where verification adds a control step, and Section 3.3 describes how candidate budgets are calibrated to meet the deadline.

We hope these additional clarifications address the remaining concerns.

Thank you for the clarification. I recognize the real-robot improvements, but several concerns remain.

  1. Evidence on large VLAs remains limited. The detailed selection analysis uses the simpler Kinetix policy at d = 4. On LIBERO, improvements with π0.5 and GR00T are more modest, around 2–5.2 percentage points at d = 2–4. The Kinetix analysis therefore does not fully explain the benefit for large VLAs.
  2. Reactive generation may already address this problem. If a fast action policy uses updated observations to generate actions in real time without waiting for the large VLA, it is unclear why selecting from stale candidates is additionally needed. These approaches may be alternatives, and calling them complementary does not establish the benefit of combining them. Although PACT-VLA avoids additional training, its added value over an existing reactive asynchronous policy remains unclear.
  3. The computational tradeoff remains unclear on constrained hardware. Table 11 obtains its largest improvement by applying verification at every decision while keeping d = 4 fixed. However, more frequent verification requires more computation. Since the latency measurements cover only powerful GPUs, it remains unclear whether these large gains would hold on constrained hardware, where verification may add control-step delay and reduce or outweigh the improvement.

Hence, I maintain my original score.

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

This paper studies the stale observation problem in asynchronous VLA control. PACT-VLA samples several action chunks at inference start, and then uses a newer observation received before the chunk switch to pick one of them. Candidates are scored by FlowRes with the frozen policy. Experiments on Kinetix, LIBERO and three real-robot tasks show clear gains.

Critical strengths
  1. The idea is simple and training-free. It uses a newer observation to pick one full chunk before execution, and the frozen policy itself works as the verifier, so no extra model is needed.
  2. The results are promising. At d=4, PACT-VLA improves over RTC by 15.0 points on Kinetix and over single-candidate GR00T by 5.2 points on LIBERO. Real-robot success goes from 60% to 82% with the same frozen checkpoint.
Critical weaknesses

The latency part needs more evaluation. Table 4 only reports N=4 at 30 Hz, while LIBERO uses N=8. On RTX 5090, latency increases from 35.4 to 60.2 ms, and both fall into d=2 at 30 Hz. But at 25 Hz, the same numbers would give d=1 and d=2. I suggest reporting the N=8 latency and explaining what happens if selection misses the deadline. This matters because, comparing PACT-VLA at d+1 with single-candidate execution at d in Table 3, one extra step of delay could lower the average LIBERO success by 4.4–13.2 points.

Questions for authors

Could the authors clarify how the model-specific settings in Table 6 (flow times, scored positions, weights and dimensions) were selected? A short sensitivity analysis would be helpful.

1 comment

We thank the reviewer for the thoughtful feedback and for recognizing the simplicity of our training-free design, the use of the frozen policy as a verifier, and the promising results across simulation and real-world tasks. We have highlighted the revisions in blue and hope our responses address your concerns.

Comment/Question:

The latency part needs more evaluation ... I suggest reporting the N=8 latency and explaining what happens if selection misses the deadline ...

Answer: We have extended the latency evaluation to N=8 and clarified how P99-based calibration and deadline-aware selection preserve the scheduled control-step delay. At 30 Hz, N=4 maintains d=2, the same as single-candidate generation, across all three evaluated GPUs (Table 4). At N=8, the delay increases to d=3 on RTX 5090 and RTX Pro 6000, while H100 remains at d=2. Since an additional control-step delay can outweigh the benefit of selection (Table 3), we calibrate N in advance using P99 latency measurements. In rare cases where verification is not complete by the commitment deadline, we select a candidate randomly from the generated pool and commit it on schedule, avoiding an additional delay. We have clarified this deadline-miss handling mechanism in Section 3.3. On LIBERO with π0.5, N=4 retains nearly all of the N=8 gain (+3.2 vs. +3.5 points at d=4, Appendix I). We use N=4 for real-robot deployment. At lower control frequencies, multi-candidate verification may incur an additional control-step delay relative to single-candidate generation (Appendix K, Table 15).

Comment/Question:

Could the authors clarify how the model-specific settings in Table 6 (flow times, scored positions, weights and dimensions) were selected? A short sensitivity analysis would be helpful.

Answer: FlowRes uses simple backbone-specific settings, with gains robust to alternative configurations (Appendix F, Fig. 7). Kinetix uses preset flow times and temporal weights without evaluation-task tuning. π0.5 inherits these flow times and uses uniform weights over executed actions, chosen after one early comparison with decaying weights and kept fixed across all suites and delays. For GR00T, we adopted a four-point flow-time grid shifted by 0.1 after a single early check and kept it fixed across all suites and delays. For π0.5 and GR00T, the action-dimension masks exclude padded coordinates, while Kinetix scores all action dimensions. Appendix F shows that all tested configurations outperform their no-selection baselines, with several alternatives even outperforming the reported π0.5 setting. Thus, the gains do not depend on finely tuned parameters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.