acceptodds
Under review as a conference paper at ICLR 2027

Looking Back to Move Forward: Temporal Verification for Generative Robot Policies

Abstract

Generative policies have emerged as a promising paradigm for robot learning, combining expressive generative action modeling with scalable imitation learning from large demonstration corpora. Yet when trained on heterogeneous demonstrations, these policies can generate suboptimal action chunks that compound errors over time, eventually driving the robot into out-of-distribution states from which recovery is difficult. Action verification offers a test-time scaling strategy for mitigating this failure mode by sampling multiple candidate actions and using a verifier to select one for execution. Existing approaches, however, remain temporally myopic and costly to train, evaluating candidates from the current observation alone without accounting for trajectory continuity and often relying on large verifiers and additional expert demonstrations. In this paper, we introduce Temporal Verification (TeV), an efficient temporally aware action verification framework for flow-matching VLAs. TeV first learns a temporal token that summarizes recent observation–action history, enabling candidate chunks to be evaluated as continuations of the execution trajectory rather than as isolated predictions. Conditioned on this token, TeV constructs positive–negative pairs without additional expert demonstrations or preference annotations and trains an energy-based verifier contrastively to assign lower energy to higher-quality, trajectory-consistent action chunks. Beyond post-hoc ranking, TeV further uses the learned energy landscape to guide intermediate flow samples toward lower-energy regions, improving candidates before final selection. Extensive experiments in simulation and real-world settings demonstrate that TeV provides reliably ranks action candidates, improves task success rates, and produces smoother execution trajectories.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot learning, Verification, Temporal consistency

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TxWCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed a method for enforcing temporal consistency by guiding action sampling using energy gradient and also using energy for selecting the optimal action from a set of candidates as an inference compute scaling approach. The paper tests the algorithm on several control tasks in both simulation and real world.

Critical strengths

The greatest strength of this paper, in my opinion, is that it demonstrates a principled way of integrating verification into the action generation process itself such that temporal consistency is bolstered. The two crucial ingredients are: (1) temporal token and (2) the energy gradient. While I find it still, somewhat, dubious that the temporal token can be trained sufficiently accurately, the paper proposes a principled manner in which the energy gradient can be used during action chunk sampling as well as post-hoc action selection. This is strictly stronger than the naive approach the reader could have considered which is simply using the verifier at test-time.

The paper carries out a meticulous set of ablations to understand the strengths and weaknesses of each component. The part that I found particularly helpful is understanding whether the energy framework is being used in a guidance-only versus verification-only setting.

The paper showcases the strengths of the method on both simulation and real world tasks that are sufficiently difficult in my opinion. In particular, the water filling task is a somewhat more dynamic task, which is interesting. The results quantifying jerkiness in figure 6 are quite illuminating.

Critical weaknesses

Minor:

  1. “Yet existing verification-based methods face two key limitations. First, they suffer from temporal myopia, generating and evaluating candidate actions based solely on the current observation. For generative policies, successful execution requires not only selecting high-quality candidates for the current observation, but also ensuring that consecutive action chunks form a coherent trajectory.” This part is misleading because both BID [1] and RTC [2] take into consideration temporal consistency. In particular, BID considers consistency with prior actions and also allowing for sufficient reactivity to unexpected changes. It is somewhat confusing because your introduction does cite these papers and, yet, says that existing methods fail on this front. If the claim is rather that these existing methods do not properly address such problems, then you should make that claim instead and provide sufficient justification for it.

  2. The specificities of what g(⋅)g(\cdot) is in equation (4) is unclear to me. Is this just the base pre-trained model? If not, what is it? How do you ensure that the temporal token is not excessively constrained by this “lightweight prediction model” g(⋅)g(\cdot)?

Major:

  1. Comparison to RTC is lacking. RTC [2] is known to be a very strong baseline and the lack of comparison to it in any setting is disappointing.
  2. All simulation testbeds are perfectly static environments. The authors can consider more stochastic environments such as multi-agent settings or even by simply adding stochastic noise to next state transitions in the static environments (similar to the experiments in BID).
  3. The paper observed that the gradient-based steering is unreliable without final candidate selection. It would be helpful to understand what exactly is the reason for it. In particular, I am interested in knowing whether this is due to the temporal token itself being sub-optimal or whether this is a modeling problem in EθE_\theta.
Questions for authors
  1. “Task success does not increase monotonically with the number of sampled candidates, indicating that additional samples are not automatically beneficial.” It would be interesting to test an alternative distribution-aware method of selectin m∗m^* (as opposed to equation 8 which simply chooses the m that minimizes the energy). As such, out-of-distribution chunks could be ruled out.
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1)

Thank you for your thoughtful and constructive feedback! In response, we have conducted additional experiments and analyses, expanded our baseline comparisons, and clarified our methodology and contributions in the revised manuscript. We hope these additions address your concerns and welcome any further discussion.

Major W1. Additional comparison with RTC

Thank you for this suggestion. To address this concern, we additionally evaluated RTC on the real-world Water Filling task, where stable and temporally consistent execution is particularly important, using the same policy backbone and evaluation protocol as our other baselines. RTC achieves a mean task-completion score of 48.8, compared with 40.5 for the base policy, 51.3 for BID, and 68.0 for TeV, representing a 19.2-point improvement over RTC.

While RTC employs training-free inpainting to maintain consistency with previously committed actions, TeV learns a history-conditioned energy verifier to evaluate and select candidate actions based on recent observation–action history. This additional comparison demonstrates TeV's stronger performance in the evaluated real-world setting, providing further evidence for the effectiveness of temporally aware action verification.

Major W2. Robustness under stochastic execution disturbances

Thank you for this valuable suggestion. To evaluate robustness beyond deterministic execution, we conducted additional experiments on RoboTwin 2.0 by injecting stochastic disturbances into executed actions, following the disturbance-testing approach in BID. Specifically, we apply temporally correlated Gaussian noise (ρ=0.8\rho=0.8) to arm-joint targets at each control step, with three disturbance levels: no noise, moderate noise (σ=0.025\sigma=0.025 radians), and strong noise (σ=0.075\sigma=0.075 radians). Perturbed commands are clipped to valid joint limits, while gripper commands remain unchanged. These disturbances introduce stochastic variations into subsequent robot states, allowing us to evaluate performance under deviations from nominal execution. We compare Base, BID, and TeV under identical disturbance settings, with 100 trials per task, method, and noise level across all seven tasks.

Table R1. Performance under stochastic execution disturbances on RoboTwin 2.0 (average success rate, %).

Noise levelBaseBIDTeV
None72.374.981.3
Moderate43.745.649.6
Strong18.617.020.3

As shown in Table R1, TeV consistently achieves the highest average success rate across all disturbance levels. Under moderate noise, TeV achieves 49.6%, outperforming Base (43.7%) and BID (45.6%). Under strong noise, all methods experience substantial degradation, with TeV achieving 20.3% compared with 18.6% and 17.0% for Base and BID. Although the performance margin narrows as disturbances intensify, these results indicate that TeV's benefits extend beyond deterministic execution. We have incorporated these additional experiments into Appendix D.

AuthorsOpenReview ↗

Rebuttal by Authors (2)

Major W3. Understanding the limitations of gradient-based guidance

Thank you for this insightful question. To investigate why gradient-based guidance is unreliable without final verification, we conducted additional experiments on LIBERO-Plus examining energy-gradient behavior, temporal conditioning, and guidance magnitude. Our findings suggest that the limitation is more closely related to the reliability of local energy-based refinement than to temporal conditioning alone. Specifically, contrastive training directly supervises relative rankings between expert chunks and constructed negatives, whereas guidance additionally requires local energy gradients to provide beneficial refinement directions during flow integration.

(1) Energy-gradient analysis. We first examine whether energy-reducing updates also improve action quality, evaluated through distance to the expert action. Across 300 held-out episodes, gradient updates reduce the learned energy in 100.0% of history-conditioned interventions and 99.1% of current-only interventions. However, among these energy-reducing updates, the distance to the expert action increases in 50.5% and 49.1% of cases, respectively. These results indicate that reducing the learned energy does not necessarily improve alignment with the expert action at each refinement step. Importantly, this diagnostic examines individual gradient updates rather than the quality of completed guided candidates, which can still benefit downstream performance when combined with final verification.

(2) Effect of temporal conditioning. To determine whether this limitation originates from the temporal token, we additionally train a current-observation-only verifier that excludes preceding observations and action history during both training and inference. Table R2 compares current-only and history-conditioned verification, with and without guidance, under the same candidate budget (M=4M=4).

Table R2. Effect of temporal history on verification and guidance (success rate, %).

BenchmarkBaseVerification (current)Verification (history)Guidance + Verification (current)Guidance + Verification (history)
LIBERO-Plus73.278.079.278.779.8
Adjust Bottle86.084.091.092.095.0
Pick Bottles48.049.062.051.064.0

History-conditioned verification outperforms current-only verification across all three settings, with these advantages persisting when guidance is enabled. We further evaluate guidance-only execution (M=1M=1) on LIBERO-Plus, where current-only and history-conditioned models achieve 56.2% and 55.9% success, respectively. Thus, removing temporal conditioning does not resolve the limitations of guidance-only execution, while incorporating history improves candidate evaluation.

(3) Effect of guidance magnitude. We also investigate whether excessive guidance strength contributes to the observed limitations. Varying the guidance cap from 0.0025 to 0.0075 yields success rates of 55.9–57.0%, while reducing guidance strength to 0.2 achieves 56.3%, all close to the default guidance-only result of 55.9%. Thus, simply weakening guidance does not consistently improve performance.

Together, these findings suggest that the limitations of guidance are more closely associated with the reliability of local energy-based refinement than with temporal conditioning or guidance magnitude alone. While EθE_\theta provides useful relative rankings for candidate selection, its local gradients do not always improve expert alignment. Final verification complements guidance by selecting among completed candidates, allowing the policy to benefit from guided generation without relying solely on individual refinement directions.

Minor W1. Distinction from BID and RTC

Thank you for identifying this ambiguity. We agree that BID and RTC explicitly address temporal consistency, and our original statement was overly broad. Our intention was to highlight the limitations of existing verification-based methods that evaluate candidates using current-step information, such as TACO, rather than to suggest that prior decoding and execution methods overlook temporal consistency.

We have revised the introduction to distinguish these approaches more clearly. Specifically, BID and RTC promote temporal consistency through their decoding and execution mechanisms, with BID also balancing continuity against responsiveness to new observations. In contrast, TeV learns a history-conditioned energy verifier that evaluates candidate actions using recent observation–action history, incorporating both action plausibility and temporal compatibility into candidate selection. Our contribution is therefore not the introduction of temporal consistency itself, but its integration into learned action verification and generation-time guidance. We have clarified this distinction and removed the overly general claim from the introduction.

AuthorsOpenReview ↗

Rebuttal by Authors (3)

Minor W2. Clarification of the auxiliary prediction head

Thank you for this question. The function g(⋅)g(\cdot) in Eq. 4 is a separate two-layer MLP prediction head, rather than the frozen pretrained VLA policy. It maps the temporal token to a prediction of the next FF actions and is trained jointly with the temporal encoder and energy verifier. This head is used only during training and discarded at inference time.

Regarding the potential constraint on the temporal token, future-action prediction serves as an auxiliary training objective alongside the contrastive verification loss, as defined in Eq. 6. The temporal token is therefore optimized not only for future-action prediction but also for distinguishing expert actions from negative candidates. Moreover, the token directly conditions the energy verifier without passing through g(⋅)g(\cdot). Thus, the auxiliary prediction head does not impose an architectural bottleneck on the information available for verification.

Q1. Distribution-aware candidate selection

Thank you for this insightful suggestion. To investigate whether incorporating distributional support improves candidate selection, we additionally evaluate a support-gated energy selection strategy while keeping the pretrained policy and verifier frozen. Specifically, we construct a reference bank of action representations sampled from the frozen policy at demonstration observations. For each candidate, we estimate its empirical support using the distance to its fifth-nearest neighbor in this bank, where smaller distances indicate greater support. We retain the 75% most-supported candidates before applying minimum-energy selection. This conservative threshold removes the lowest-support candidate(s) while preserving alternatives for energy-based ranking (retaining 3, 6, and 12 candidates for M=4,8,16M=4,8,16, respectively), without additional model training.

Table R3. Effect of distribution-aware candidate selection on RoboTwin 2.0 (success rate, %).

TaskCandidates (M)Energy-OnlySupport-Gated EnergyΔ (pp)
Adjust Bottle495.093.0−2.0
888.087.0−1.0
1687.087.00.0
Press Stapler454.054.00.0
857.056.0−1.0
1658.058.00.0

As shown in Table R3, support-gated selection does not outperform minimum-energy selection at any evaluated candidate budget, nor does it mitigate the performance decline on Adjust Bottle as the candidate count increases. One possible explanation is that feature-space proximity to the reference distribution does not necessarily reflect action quality under the current execution context. These results suggest that additional support-based filtering, under the evaluated configuration, provides no improvement over history-conditioned energy verification, supporting our use of minimum-energy selection.

Dear Reviewer TxWC,

Thank you again for your time and thoughtful feedback on our paper! As the author–reviewer discussion period draws to a close, we wanted to follow up and express our hope that the additional evidence and revisions have addressed your concerns. Your suggestions have helped us to strengthen the paper, and we would be happy to discuss any further questions.

Best regards,

The Authors

Reviewer_9k64OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a technique to improve verification-based action selection during deployment time, by both scoring samples and guiding generation using an energy model. The key is the encoding of recent observation and action history into a temporal token, which is provided as conditioning to a learned energy-based verification model. The energy-based verification model is trained to assign lower energy (higher likelihood) to action candidates that are more likely to occur after the historical context. The proposed approach can be integrated cheaply with existing VLAs.

Critical strengths

This paper is clearly written and the approach intuitively makes sense to try. The experiments also appear quite complete, including real-world experiments. Thorough ablations over the most relevant components of the method (negative example strategy, roles of guidance and verification) are provided.

Critical weaknesses

One critical weakness is the lack of explanation for how alternate approaches are trained, to ensure fair comparison. For example, there are no details on how the dynamics models of LPB or DynaGuide or FOREWARN are trained, for Table 1. Would they use the same amount of data as used in TeV, to keep evaluations fair? Without these details, it is hard to extract takeaways from the experimental results.

The variance of the real-world experiments also seems quite high (Table 4), making comparison with base or alternate approaches (like TACO) less clear.

The energy score reliability also seems lower than expected, for a model that has been trained contrastively (Figure 5). It does seem like there are additional ways to construct negatives to increase energy score reliability (beyond 50% for the best result, which seems like a coin flip...).

In terms of related work, the approach of generating multiple candidates (including with early pruning, as described in Appendix B as "progressive pruning" [there is also a typo at the end of that paragraph]) and then ranking the final approach has been done in text-to-image generation [1]. The difference is that SGI does not steer generation, but the general approach of pruning and scoring is similar - it would be interesting if the TeV authors also try multiple rounds of pruning as well. Another related thread of work is using Best-of-N selection for sample-efficient self-improvement [2]. Currently, TeV only does Best-of-N selection for one round of inference - it would be interesting to provide additional results to see if such benefits from TeV also extend to self-improvement over iterations compared to alternate approaches. As TeV relies on Best-of-N action selection, it would also make sense to reference relevant past works that propose and use Best-of-N [3, 4].

[1] Parmar et al., Scaling Group Inference for Diverse and High-Quality Generation, ICLR 2026.

[2] Luo et al., DF-ExpEnse: Diffusion Filtered Exploration for Sample Efficient Finetuning, ICML 2026.

[3] Ghasemipour et al., EMaQ: Expected-Max Q-Learning Operator for Simple Yet Effective Offline and Online RL, ICML 2021.

[4] Li et al., Reinforcement Learning with Action Chunking, NeurIPS 2025.

Questions for authors

Why are baseline methods for LIBERO in Table 1 not provided in RoboTwin2.0 for Table 2?

Are there any other negative example construction strategies that the authors tried (but found didn't work)? Any other strategies that the authors think are worth trying, or are the two presented considered sufficient?

What were the sizes of KmisK_{mis} and KcrsK_{crs}? Would we be able to see ablations over their sizes to show if one negative strategy is more helpful than another? How many negatives as a whole are needed to learn a good energy model contrastively?

6 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1)

Thank you for your thoughtful and constructive feedback! In response, we have conducted additional ablation experiments and analyses, expanded baseline comparisons, clarified implementation details, and revised the manuscript accordingly. We hope these additions address your concerns and welcome any further discussion.

W1. Baseline training fairness and implementation details

Thank you for raising this important concern. To ensure a fair comparison, TeV and all learned baselines are trained using the same corpus of 1,693 LIBERO expert demonstrations and the same frozen π0.5\pi_{0.5} policy checkpoint. No additional environment rollouts or human annotations are used to train these auxiliary models. Below, we clarify the training and implementation details for each baseline.

LPB and DynaGuide. We implement a shared action-conditioned latent dynamics model using a frozen DINOv2 visual encoder. The model predicts future observation embeddings over a 10-step horizon and is trained with an MSE objective on demonstration trajectories, using both clean and flow-noised action inputs. DynaGuide uses the learned dynamics for gradient-based action steering, whereas LPB additionally applies a distance-based intervention criterion. Both methods therefore use the same training data and underlying dynamics model, with differences arising from their respective intervention strategies.

V-GPS. We train twin Q-networks and a state-value network using temporal-difference learning. Reward labels are derived automatically from successful demonstrations. The learned Q-networks rank candidate chunks using the minimum twin-Q estimate of their first actions. Thus, V-GPS uses demonstration-derived reward supervision without additional interactions or annotations.

TACO. We train a Coin-Flipping Network on cached expert action-chunk representations from the same policy and demonstrations. The network regresses randomly assigned {-1,+1} targets using MSE, and its inverse squared output norm provides a pseudo-count score for candidate selection.

FOREWARN. The original implementation additionally requires policy rollouts and an 11B VLM alignment stage with annotated behavior narrations. To maintain the same demonstration-only training setting, we adapt its outcome-based candidate selection approach using a lightweight latent dynamics model that predicts candidate outcomes and ranks them against demonstration-derived references. Since this differs from the original implementation, we explicitly denote it as FOREWARN$^\dagger$ to distinguish the evaluated variant from the full method. We have added these training and implementation details to the revised manuscript, clarifying the shared data budget and the scope of each baseline adaptation.

W2. Variance and reliability of real-world performance

Table 4 reports task-completion scores rather than binary success rates, with each trial evaluated on a 0–100 scale based on completed task stages (Appendix B). The relatively large standard deviations reflect variation in how far individual executions progress before failure. Importantly, TeV achieves both higher mean completion scores and lower standard deviations than all compared baselines across the three tasks. For example, on Water Filling, TeV achieves 68.0 ± 12.8, compared with 51.5 ± 29.7 for TACO and 40.5 ± 32.4 for the base policy.

To further characterize the performance distributions, we added performance profiles [1] in Appendix D, Figure 8. These profiles report the fraction of trials achieving at least a given completion score, allowing us to examine performance across different task-progress thresholds. TeV outperforms the baselines across nearly all thresholds, confirming that its performance gains extend across the outcome distribution rather than being driven by a few high-scoring trials. These findings further reinforce the performance advantages observed in Table 4, supporting the consistency of TeV's improvements across real-world executions.

[1] Deep reinforcement learning at the edge of the statistical precipice. NeurIPS 2021.

AuthorsOpenReview ↗

Rebuttal by Authors (2)

W3. Energy-ranking reliability

Thank you for this concern. We first clarify that Figure 5 evaluates rankings among four candidate action chunks, where random selection has a 25% probability of identifying the candidate closest to the expert representation, rather than 50%. Across the three real-world tasks, the verifier achieves 48.7–51.5% top-1 agreement and selects one of the two nearest candidates in 73.5–75.5% of states, compared with random baselines of 25% and 50%, respectively. Thus, the verifier identifies the closest candidate approximately twice as often as random selection, demonstrating meaningful discrimination above chance.

Importantly, our contrastive objective learns relative compatibility between expert actions and constructed negatives, rather than explicitly optimizing the latent-distance ordering of generated candidates. Figure 5 measures agreement with this distance-based ranking, which does not necessarily reflect action quality, as multiple valid continuations may differ from a single expert reference. Therefore, imperfect ranking agreement does not imply ineffective verification, as further supported by the downstream performance improvements in Tables 1, 2, and 4.

Regarding negative construction, our two negative families target complementary aspects of action quality: temporal compatibility and policy-generated deviations. We further investigate alternative negative constructions in our response to Q2.

W4. Relation to Best-of-N selection, progressive pruning, and iterative self-improvement

Thank you for identifying these relevant works and suggesting additional experiments. We agree that Best-of-N selection and progressive pruning are established techniques. Our contribution lies in learning a history-conditioned energy verifier for both candidate selection and generation-time guidance, rather than introducing these sampling strategies. We have incorporated the suggested references [2–5] and clarified this distinction.

Progressive pruning and SGI. We clarify that the main RoboTwin 2.0 results in Table 2 use no pruning, whereas progressive pruning is employed in our real-world experiments to reduce inference latency. Motivated by SGI [2], we additionally investigate whether multiple pruning stages improve TeV's performance. We compare single-stage pruning (4→2 candidates after integration step 5) with two-stage pruning (4→3 after step 3, then 3→2 after step 7). Both pruning variants use 30 candidate-level integration steps, compared with 40 without pruning.

Table R1. Effect of progressive pruning on RoboTwin 2.0 (success rate, %).

Pruning strategyAdjust BottlePick Bottles
No pruning95.064.0
Single-stage95.068.0
Two-stage95.069.0

Table R1 shows both pruning strategies maintain 95.0% success on Adjust Bottle while improving Pick Bottles from 64.0% without pruning to 68.0% and 69.0%, respectively. The two pruning schedules achieve comparable performance, indicating that intermediate pruning can reduce sampling computation while maintaining or improving task success.

Best-of-N and iterative self-improvement. We additionally investigate whether TeV-based candidate selection can improve subsequent policy learning. Following a candidate-selection-based collection and policy-update framework [3], we compare ordinary DSRL-SAC collection [6], DF-ExpEnse-style quality-and-uncertainty selection, and TeV minimum-energy selection. All methods use the same frozen π0.5\pi_{0.5} backbone and train a noise policy over two rounds, collecting 20,000 executed control steps per round. After each update, we evaluate the learned policy using a single action sample without test-time selection, isolating improvements retained through policy learning.

Table R2. Iterative policy improvement on RoboTwin 2.0 (success rate, %).

TaskStrategyRound 0Round 1Round 2
Adjust BottleDSRL-SAC86.082.082.0
DF-ExpEnse-style86.088.082.0
TeV min energy86.082.085.0
Pick BottlesDSRL-SAC48.037.048.0
DF-ExpEnse-style48.057.049.0
TeV min energy48.041.045.0

Table R2 shows that neither TeV-based nor DF-ExpEnse-style collection produces consistent improvements across successive policy updates. These results indicate that effective inference-time candidate selection does not automatically translate into sustained improvements through iterative policy learning. We have included these findings in Appendix D.

[2] Scaling Group Inference for Diverse and High-Quality Generation.

[3] DF-ExpEnse: Diffusion Filtered Exploration for Sample Efficient Finetuning.

[4] EMaQ: Expected-Max Q-Learning Operator for Simple Yet Effective Offline and Online RL.

[5] Reinforcement Learning with Action Chunking.

[6] Steering Your Diffusion Policy with Latent Space Reinforcement Learning.

AuthorsOpenReview ↗

Rebuttal by Authors (3)

Q1. Additional baseline comparisons on RoboTwin 2.0

Thank you for pointing this out. Our original RoboTwin 2.0 evaluation focused on methods most directly related to temporal consistency and candidate verification (TE, BID, and TACO), while broader dynamics-based approaches were evaluated on LIBERO-Plus. Among these additional methods, FOREWARN$^\dagger$ achieved the strongest performance on LIBERO-Plus (75.0%), compared with V-GPS (67.2%), LPB (58.5%), and DynaGuide (58.6%). We therefore prioritized FOREWARN$^\dagger$ for additional evaluation on RoboTwin 2.0 as the most competitive baseline from this group.

Table R3. Additional RoboTwin 2.0 comparisons (success rate, %).

MethodAdjust BottlePick BottlesPlace ContainerStack BowlsPlace CupOpen LaptopPress StaplerAvg
Base86.048.086.089.078.075.044.072.3
FOREWARN$^\dagger$87.054.086.091.080.062.047.072.4
TeV95.064.094.093.088.081.054.081.3

As shown in Table R3, FOREWARN$^\dagger$ achieves 72.4% average success, comparable to the base policy (72.3%), whereas TeV achieves 81.3%, outperforming FOREWARN$^\dagger$ by 8.9 percentage points. This additional comparison extends our evaluation to the strongest of the previously omitted baselines on LIBERO-Plus, providing further evidence of TeV's effectiveness across benchmarks. We have incorporated these results into the revised manuscript.

Q2. Alternative negative construction strategies

Thank you for this suggestion. Beyond the proposed temporal-mismatch and coarse-integration negatives, we additionally evaluated Gaussian-perturbed expert chunks as an alternative negative construction. Specifically, we add Gaussian noise ϵ∼N(0,σ2I)\epsilon \sim \mathcal{N}(0,\sigma^2 \mathbf{I}) to expert action chunks, testing σ∈\sigma \in {0.1, 0.3}. To isolate the effect of this alternative, we replace coarse-integration negatives with Gaussian-perturbed negatives while retaining temporal-mismatch negatives.

Table R4. Performance (%) of different negative construction strategies on RoboTwin 2.0.

MethodAdjust BottlePick BottlesPlace ContainerStack BowlsPlace CupOpen LaptopPress StaplerAvg
Base86.048.086.089.078.075.044.072.3
Coarse Only89.062.089.091.085.079.052.078.1
Temporal Only91.059.092.091.084.076.049.077.4
Coarse + Temporal (Ours)95.064.094.093.088.081.054.081.3
Gaussian Jitter (σ = 0.1) + Temporal89.061.090.089.084.079.047.077.0
Gaussian Jitter (σ = 0.3) + Temporal90.059.090.088.083.077.048.076.4

As shown in Table R4, Gaussian-perturbed negatives achieve 77.0% and 76.4% average success, respectively, providing no improvement over temporal-mismatch negatives alone (77.4%). In contrast, combining coarse-integration and temporal-mismatch negatives achieves 81.3%, outperforming both Gaussian variants by 4.3–4.9 percentage points. These results suggest that policy-generated deviations provide a more effective training signal than generic Gaussian perturbations in our setting.

While we do not claim that the two negative families are exhaustive, their complementary design addresses two distinct failure modes: temporal incompatibility and suboptimal policy-generated actions. Their combination outperforms either family alone across all seven tasks, providing empirical support for the sufficiency of the proposed construction in our evaluated settings without requiring additional demonstrations or preference annotations. We have included these additional ablations in the revised manuscript.

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

This paper proposes a history-conditioned energy-based verifier for a frozen flow-matching VLA (π0.5 used in the paper). The verifier is trained with a soft margin contrastive loss and used to rank candidate action chunks generated from the VLA, and optionally guide generation. Experiments show that the proposed method achieves competitive performance on sim and real benchmarks.

Critical strengths
  • The idea generally makes sense and experiments show improvements across sim and real benchmarks, with the same VLA backbone for comparison.
  • The method reuses existing demonstrations and trains relatively small additional modules, which might be a decent contribution to efficient downstream fine-tuning.
Critical weaknesses
  • Additional ablations on the main method. The main claims are twofold. One is history; two is verification. After reading the manuscript and the ablations, my informed guess is that simply learning the verifier without guidance or history might be a good enough downstream fine-tuning step in this setup. To test verification without history, I think we are missing a current-frame verifier ablation here. To test adding context without the verifier, I am unsure if there is a way to do this without unfreezing the policy, so I leave this as a question to the authors. On balance, it would strengthen main claim if we could have a current observation-only verifier baseline.
Questions for authors
  • Do you see guidance helping the policy performance?
  • Additional ablations: see weaknesses.
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for the thoughtful and constructive feedback! In response, we have conducted additional controlled ablations and analyses to strengthen the empirical evidence and clarify our methodological contributions. We hope these additions address your concerns and welcome any further discussion.

W1. Additional ablations isolating the contribution of temporal history

Thank you for this valuable suggestion. To disentangle the contributions of verifier learning and temporal history, we conducted additional controlled ablations comparing current-observation-only and history-conditioned verification, both with and without gradient-based guidance (Table R1). The current-only variant excludes preceding observations and executed actions during both training and inference. Both variants use the same frozen base policy, negative construction, and training protocol, isolating the effect of temporal conditioning. We evaluate LIBERO-Plus and two RoboTwin 2.0 tasks spanning different baseline performance levels.

Table R1. Effect of temporal history on verification and guidance (success rate, %).

BenchmarkBaseVerification (current)Verification (history)Guidance + Verification (current)Guidance + Verification (history)
LIBERO-Plus73.278.079.278.779.8
Adjust Bottle86.084.091.092.095.0
Pick Bottles48.049.062.051.064.0

The results reveal that the contribution of temporal history varies across benchmarks. On LIBERO-Plus, current-only verification already improves success from 73.2% to 78.0%, with temporal history providing an additional 1.2 percentage points. In contrast, on RoboTwin 2.0, current-only verification performs comparably to the base policy (84.0% vs. 86.0% on Adjust Bottle; 49.0% vs. 48.0% on Pick Bottles), whereas history-conditioned verification achieves 91.0% and 62.0%, respectively, improving over current-only verification by 7.0 and 13.0 points. These advantages persist when guidance is enabled. Thus, while learning a verifier alone accounts for much of the improvement on LIBERO-Plus, temporal history provides substantial additional benefits on the evaluated RoboTwin tasks, independently of gradient-based guidance.

Regarding history without verification, TeV introduces temporal context through the verifier while keeping the base policy frozen. Removing the verifier would therefore also remove the mechanism through which history influences action selection. Evaluating history alone would require a separate history-based selection mechanism or modifying the policy's conditioning. As complementary comparisons, the existing TE and BID baselines incorporate information from previous action predictions without a learned verifier. We have added these results and analysis in the revised manuscript.

Q1. Does guidance improve policy performance?

Yes. Guidance provides additional performance gains when combined with verification, although its effectiveness varies across benchmarks. As shown in Figure 3, verification alone achieves 72.7% on the evaluated RoboTwin 2.0 tasks, while adding guidance improves performance to 75.1–77.3% across the tested guidance strengths, achieving up to a 4.6-percentage-point improvement. On LIBERO-Plus, the additional benefit is smaller, with guidance achieving 78.8–79.9% compared with 79.2% for verification alone.

These results reflect the complementary roles of the two components: guidance refines intermediate flow samples, while verification selects the lowest-energy completed action chunk. Although guidance alone is less reliable, combining it with final verification allows the policy to benefit from candidate refinement while selecting among the resulting actions. Thus, guidance provides a measurable additional benefit, while verification remains the primary source of improvement.

Dear Reviewer efzX,

Thank you again for your time and thoughtful feedback on our paper! As the author–reviewer discussion period draws to a close, we hope that our additional experiments and revisions have addressed your concerns. Your suggestions have helped strengthen the paper, and we would be happy to discuss any further questions.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Follow-up: Summary of Revisions

We sincerely thank all reviewers for their thoughtful and constructive feedback. We particularly appreciate their recognition of the principled integration of temporal verification into action generation (TxWC), the efficient and lightweight design leveraging existing demonstrations (efzX), and the extensive empirical validation across simulation and real-world tasks, including informative component ablations (efzX, 9k64, TxWC).

In response to the reviewers' concerns, we have strengthened the manuscript through methodological clarifications, additional experiments, and deeper analyses. These revisions provide further empirical support for TeV's central contribution: a lightweight, history-conditioned energy verifier that enables both candidate selection and generation-time guidance without modifying the pretrained VLA policy. We summarize the major revisions below.

1. Additional ablations isolating the contribution of temporal history. We introduced current-observation-only verifier baselines, with and without guidance, to isolate the contribution of temporal history from verifier learning. The results demonstrate consistent additional gains from temporal conditioning across the evaluated benchmarks, with varying magnitudes, directly validating its contribution independently of gradient-based guidance.

2. Baseline clarification and expanded experimental validation. We clarified the training protocols and supervision requirements of the compared baselines, extended the comparisons to include dynamics-based verification and RTC, and evaluated robustness under stochastic execution disturbances. We also added real-world progress profiles to better characterize performance variability.

3. Deeper analysis of verifier learning and guidance. We expanded the energy-ranking reliability analysis, investigated alternative negative constructions and sampling budgets, and examined the effects of temporal conditioning, energy gradients, and guidance magnitude. These analyses provide further insights into the learned verifier and clarify the complementary roles of guidance and final candidate selection.

4. Extensions and methodological positioning. We clarified TeV's relationship to existing temporal-consistency and Best-of-N methods and explored its integration with multi-stage progressive pruning, distribution-aware candidate selection, and iterative policy learning. These additional experiments provide a broader understanding of TeV's capabilities while further distinguishing its methodological contributions from prior work.

We thank the reviewers again for their valuable feedback, which has helped us strengthen the work. We hope our additional results and revisions address the concerns and would be happy to address any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.