acceptodds
Under review as a conference paper at ICLR 2027

MRFT: MeanFlow Reinforcement Fine-Tuning via Tractable Stochastic Transitions

Abstract

Flow matching has shown strong performance in continuous robotic control, but its multi-step sampling procedures incur substantial rollout costs in online reinforcement learning (RL). MeanFlow presents a compelling solution to this sampling overhead by enabling efficient few-step generation through explicitly modeling the average velocity field. However, directly applying MeanFlow to online RL is hindered by a critical limitation: its deterministic sampling process lacks the stochastic exploration and tractable transition likelihoods required for RL fine-tuning. Existing techniques depend on auxiliary learnable networks for post-hoc stochasticity injection, inherently increasing computational overhead and parameter complexity. Rather than relying on heuristic designs, we formally derive MeanFlow-SDE from the underlying flow dynamics. By approximating the finite-interval integration of the marginal-preserving flow SDE, MeanFlow-SDE yields tractable Gaussian transitions that inherently support stochastic exploration, eliminating the need for additional components. Building upon MeanFlow-SDE, we introduce MeanFlow Reinforcement Fine-Tuning (MRFT), an efficient online fine-tuning framework tailored for MeanFlow policies. By embedding our tractable transitions into RL objectives, MRFT seamlessly integrates stochastic generative sampling with robust policy optimization. Extensive evaluations on lightweight control policies, -style architectures, and real-world robotic systems demonstrate consistent improvements in sample efficiency and task success rates while maintaining few-step inference.

Primary area:
reinforcement learning
Keywords:
Few-Step Generative Policies, MeanFlow, Reinforcement Fine-Tuning, Robotic Manipulation

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_3JorOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper works on how to do RL on a deterministic MeanFlow. The resultant MRFT achieves better or on-par results on multiple simulations and real robot tasks.

I hope to kindly report to the AC that my expertise in RL theory is limited, which I hope the AC can take into considerations.

Critical strengths
  1. A full mathematical proof is offered to support the proposed method.
  2. Both real robot and simulation results are offered.
Critical weaknesses
  1. The proposed method seems to outperform the baseline very little in Figure 1. It would be great if more analysis can be provided to justify this, and to offer insights on where MRFT is suitable or not.
Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3Jor (Part 1)

We deeply appreciate your insightful comments and efforts in reviewing our manuscript. We respond to each of your comments one by one in what follows.
[W1]: The proposed method seems to outperform the baseline very little in Figure 1. It would be great if more analysis can be provided to justify this, and to offer insights on where MRFT is suitable or not.
[W1.1]: The proposed method seems to outperform the baseline very little in Figure 1. It would be great if more analysis can be provided to justify this.
Reply: Thank you for raising this concern. To clarify the performance differences in Figure 1, we first report the detailed results from Appendix F below, making both the magnitude of the improvements and the variability across runs explicit. Results are reported as mean ± standard deviation over three random seeds.

Gym Locomotion Tasks (Average Episode Reward)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Gym-ant-v24040 ± 82.44090 ± 14.24277±39.6‾\underline{4277 \pm 39.6}4376 ± 33.9
Gym-hopper-v23208 ± 16.13188±15.3‾\underline{3188 \pm 15.3}2942 ± 25.23074 ± 31.9
Gym-humanoid-v25198±155.3‾\underline{5198 \pm 155.3}5054 ± 49.15134 ± 134.15238 ± 13.3

Franka Kitchen Tasks (Task Completion Rate, %)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Kitchen-complete78.7 ± 2.596.3 ± 0.799.3±0.6‾\underline{99.3 \pm 0.6}99.6 ± 0.3
Kitchen-mixed74.8±0.1‾\underline{74.8 \pm 0.1}74.4 ± 0.674.2 ± 0.375.0 ± 0.0
Kitchen-partial64.9±0.1‾\underline{64.9 \pm 0.1}58.6 ± 13.850.0 ± 0.071.8 ± 5.1

Robomimic Tasks (Success Rate, %)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Robomimic-can98.7±1.5‾\underline{98.7 \pm 1.5}97.7 ± 0.698.7±1.5‾\underline{98.7 \pm 1.5}99.7 ± 0.6
Robomimic-square73.9 ± 8.375.2 ± 7.886.3±1.7‾\underline{86.3 \pm 1.7}90.1 ± 0.9
Robomimic-transport93.2 ± 1.080.3 ± 2.195.5±3.6‾\underline{95.5 \pm 3.6}97.9 ± 1.9

Bold and underlined values indicate the best and second-best results in each task, respectively.

The observed gains on Kitchen-Partial and Square are substantial rather than merely marginal: MRFT exceeds the strongest competing baseline by 6.9 and 3.8 percentage points, respectively. Overall, MRFT achieves the highest mean performance on eight of nine tasks. Starting from the same pretrained MeanFlow policy, it improves mean performance over DMPO on all nine tasks, including gains of 21.8 percentage points on Kitchen-Partial and 99–132 reward points across the three Gym tasks. We provide a more detailed analysis in Appendix F for your reference.
For context, DMPO reports a 3-percentage-point gain in peak success rate over ReinFlow-R on Square (Section 4.2.1), while D²PPO$^{[1]}$ reports a 6-percentage-point gain over DPPO$^{[2]}$ on Square after fine-tuning (Table 2). These reported gains provide context for the magnitude of our improvements.
The gains also extend to training efficiency. Under the same interaction budget and hardware, MRFT achieves the highest mean success rate across all three Robomimic tasks, with iteration times 9.1–18.3% shorter than those of DMPO and 3.3–21.0% shorter than those of ReinFlow-S, while remaining comparable to those of ReinFlow-R (Table 1 of the manuscript).
Additionally, to make the performance differences easier to discern, we have revised Figure 1 by adding zoomed-in insets to all three Gym tasks, highlighting the key regions where the differences between methods are more visible.


[1] Diffusion Policy Policy Optimization with Dispersive Loss, AAAI 2026.
[2] Diffusion policy policy optimization, ICLR 2025.

AuthorsOpenReview ↗

Response to Reviewer 3Jor (Part 2)

[W1.2] It would be great if offering insights on where MRFT is suitable or not.
Reply: We thank the reviewer for encouraging some insights on where MRFT is suitable or not. MRFT is well suited to online fine-tuning of pretrained MeanFlow policies for continuous control, enabling stochastic exploration while retaining few-step action generation. Its broader applicability comes from MeanFlow-SDE, which constructs stochastic finite-interval transitions directly from the existing MeanFlow field, requiring one model evaluation per interval without an auxiliary learnable noise network or anchor-based resampling. Within the MeanFlow formulation studied here, this construction is independent of the downstream task and policy backbone. Its tractable sampling-path likelihoods also support both on-policy and off-policy optimization (Section 4).
Our experiments support this flexibility across locomotion, state-based and visual manipulation, a π0.5\pi_{0.5}-style VLA, and online fine-tuning on a physical robot (Sections 5.1–5.4). In practice, MRFT combines controllable stochastic exploration with efficient few-step MeanFlow sampling. The noise level provides a simple way to adjust exploration strength, while our ablations show that two model evaluations per sample achieve a favorable balance between task performance and training cost (Section 5.3). These properties make MRFT particularly attractive for online reinforcement fine-tuning, where action generation is repeated throughout training.
Beyond robotics, related SDE-based sampling constructions have enabled RL fine-tuning in image generation (Flow-GRPO$^{[1]})andvideogeneration(DanceGRPO) and video generation (DanceGRPO^{[2]}$). MeanFlow-SDE offers a corresponding route to reward-based fine-tuning of conditional MeanFlow generators in these domains, although these extensions remain to be evaluated.


[1] Flow-GRPO: Training Flow Matching Models via Online RL, NIPS 2025.
[2] DanceGRPO: Unleashing GRPO on Visual Generation, Arxiv 2025.

I sincerely thank the authors for the explanations and justifications. I suggest adding them to the main text for broader readers to better understand the method's significance. I remain weak accept.

Dear Reviewer 3Jor,

Thank you very much for taking the time to consider our explanations and for your helpful suggestion!

We have moved the final performance table into the main text and incorporated the key insights from into the conclusion. We agree that these changes help readers better understand the performance and significance of our method.

If our responses have helped resolve your concerns, might you also feel comfortable revisiting your confidence level? We would, of course, fully respect your judgment.

Best regards, The Authors

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

The paper proposes an approach that reduces velocity model evaluation at each sampling step to accelerate flow-SDE sampling. Then, the work applies it to RL fine-tuning of flow policies for robotic control tasks. The core idea is to approximate the velocity integral along an SDE trajectory with an ODE trajectory and use MeanFlow to estimate the average velocity.

Critical strengths
  1. The proposed MeanFlow SDE transition is theoretically sound. The paper is able to bound the approximation error from the ODE velocity integral approximation.
  2. Experiments are designed and structured in a clear way and target good coverage of research questions and settings, including real-robot experiments.
Critical weaknesses
  1. Novelty compared to existing literature: there seem to be existing relevant works on accelerating SDE sampling via MeanFlow. Their core techniques can be similar even if they’re not applied to the RL fine-tuning setting. The paper should carefully discuss this type of works and distill the novelty. Some examples I found (non-extensive):
    1. Stochastic Transition-Map Distillation for Fast Probabilistic Inference, Rapakoulias et al., 2026.
    2. Stochastic Few-step Models, Passaro et al., 2026.
  2. Empirical performance:
    1. Performance improvement over other baselines on few-step SDE sampling is relatively marginal; also, it seems that only 1 random seed is used for experiments in section 5.2. Since the performance is so close between methods there, it is hard to say if the observed performance gain is just statistical noise or actual improvement that is statistically significant.
    2. Real-robot experiments: there is no baseline comparison here, and the final performance after fine-tuning is rather limited (around 60%-70% success rate).
  3. Reproducibility: experiment details, how hyper-parameters are tuned for the proposed approach and the baselines are not specified.
  4. Small things:
    1. Section 3 problem formulation: it seems that the paper is dealing with a finite-horizon MDP, then, the definition of the MDP tuple should include the horizon
Questions for authors

N/A

5 comments
AuthorsOpenReview ↗

Response to Reviewer oxBK (Part 1)

Thank you for your thoughtful review and insightful comments. We address each comment below.
[W1] Novelty compared to existing literature: there seem to be existing relevant works on accelerating SDE sampling via MeanFlow. Their core techniques can be similar even if they’re not applied to the RL fine-tuning setting. The paper should carefully discuss this type of works and distill the novelty. Some examples I found: Stochastic Transition-Map Distillation for Fast Probabilistic Inference, Rapakoulias et al., 2026; Stochastic Few-step Models, Passaro et al., 2026.
Reply: Thank you for pointing out these works. Our work focuses on deriving a stochastic sampling formula for MeanFlow itself to enable reinforcement fine-tuning, while these studies explore MeanFlow to accelerate SDE sampling in other models.
Specifically, the paper "Stochastic Transition-Map Distillation for Fast Probabilistic Inference" uses conditional MeanFlow as a deterministic ODE-based generator of posterior samples p(x0∣xt)p(x_0\mid x_t). This approach uses MeanFlow’s few-step capability to accelerate diffusion-model inference. The paper "Stochastic Few-step Models" constructs reparameterized conditional ODEs matching the SDE’s conditional marginals, then uses MeanFlow to accelerate these ODEs. In both cases, MeanFlow serves as a deterministic transport module within a stochastic sampling framework.
Our method directly equips a MeanFlow policy with a stochastic finite-interval sampler. We derive an approximate Gaussian update from the underlying Flow-SDE using the policy’s original average-velocity field, without a learnable noise network or a separate resampling stage. This enables stochastic exploration and tractable path likelihoods for reinforcement fine-tuning of meanflow policies.
[W2] Empirical performance
[W2.1] Performance improvement over other baselines on few-step SDE sampling is relatively marginal.
Reply: Thank you for raising this concern. We first report the detailed results from Appendix F below to make both the magnitude of the improvements and the variability across runs explicit. Results are reported as mean ± standard deviation over three random seeds.

Gym Locomotion Tasks (Average Episode Reward)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Gym-ant-v24040 ± 82.44090 ± 14.24277±39.6‾\underline{4277 \pm 39.6}4376 ± 33.9
Gym-hopper-v23208 ± 16.13188±15.3‾\underline{3188 \pm 15.3}2942 ± 25.23074 ± 31.9
Gym-humanoid-v25198±155.3‾\underline{5198 \pm 155.3}5054 ± 49.15134 ± 134.15238 ± 13.3

Franka Kitchen Tasks (Task Completion Rate, %)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Kitchen-complete78.7 ± 2.596.3 ± 0.799.3±0.6‾\underline{99.3 \pm 0.6}99.6 ± 0.3
Kitchen-mixed74.8±0.1‾\underline{74.8 \pm 0.1}74.4 ± 0.674.2 ± 0.375.0 ± 0.0
Kitchen-partial64.9±0.1‾\underline{64.9 \pm 0.1}58.6 ± 13.850.0 ± 0.071.8 ± 5.1

Robomimic Tasks (Success Rate, %)

TaskReinFlow-RReinFlow-SDMPOMRFT (ours)
Robomimic-can98.7±1.5‾\underline{98.7 \pm 1.5}97.7 ± 0.698.7±1.5‾\underline{98.7 \pm 1.5}99.7 ± 0.6
Robomimic-square73.9 ± 8.375.2 ± 7.886.3±1.7‾\underline{86.3 \pm 1.7}90.1 ± 0.9
Robomimic-transport93.2 ± 1.080.3 ± 2.195.5±3.6‾\underline{95.5 \pm 3.6}97.9 ± 1.9

Bold and underlined values indicate the best and second-best results in each task, respectively.

The observed gains on Kitchen-Partial and Square are substantial rather than merely marginal: MRFT exceeds the strongest competing baseline by 6.9 and 3.8 percentage points, respectively. Overall, MRFT achieves the highest mean performance on eight of nine tasks. Starting from the same pretrained MeanFlow policy, it improves mean performance over DMPO on all nine tasks, including gains of 21.8 percentage points on Kitchen-Partial and 99–132 reward points across the three Gym tasks.
For context, DMPO reports a 3-point gain in peak success rate over ReinFlow-R on Square (Section 4.2.1), while D²PPO$^{[1]}$ reports a 6-point gain over DPPO$^{[2]}$ after fine-tuning (Table 2). These provide a reference for the magnitude of our improvement.
The gains also extend to training efficiency. Under the same interaction budget and hardware, MRFT achieves the highest mean success rate across all three Robomimic tasks, with 9.1–18.3% shorter iteration times than DMPO and 3.3–21.0% shorter than ReinFlow-S, and times comparable to ReinFlow-R (Table 1 of the manuscript).
Additionally, we have added zoomed-in insets for all three Gym tasks in Figure 1 to highlight performance differences between methods.


[1] Diffusion Policy Policy Optimization with Dispersive Loss, AAAI 2026.
[2] Diffusion policy policy optimization, ICLR 2025.

AuthorsOpenReview ↗

Response to Reviewer oxBK (Part 2)

[W2.2] It seems that only 1 random seed is used for experiments in section 5.2. Since the performance is so close between methods there, it is hard to say if the observed performance gain is just statistical noise or actual improvement that is statistically significant.
Reply: The experiments in Section 5.2 followed the LIBERO setup and evaluation protocol of π\pi-RL$^{[1]}$. We agree with your concern and agree that multiple seeds are needed to assess variability across training runs. We therefore repeated RL fine-tuning with two additional seeds for each method, initializing all runs from the same few-shot SFT checkpoint. The updated table reports the SFT reference scores and each RL method’s success rate (%) as the mean ± std across three seeds. Δ Avg. denotes the gain in average success rate over few-shot SFT, in percentage points.

ModelMethodSpatialObjectGoalLongAvg.Δ Avg.
# Full-Data SFT
π0.5\pi_{0.5}-MeanFlowSFT98.095.496.891.495.4—
# Few-Shot SFT + RL
π0.5\pi_{0.5}-MeanFlowSFT97.483.891.089.490.4—
Flow-Noise98.1 ± 0.493.5 ± 0.198.2 ± 0.493.9 ± 0.496.0 ± 0.1+5.6 ± 0.1
MRFT (ours)98.7 ± 0.397.0 ± 0.698.5 ± 0.394.5 ± 0.597.2 ± 0.2+6.8 ± 0.2

The three-seed results show that MRFT maintains higher mean success rates than the baseline across all four suites, particularly on LIBERO-Object with gains of 3.5.
[W2.3] Real-robot experiments: there is no baseline comparison here.
Reply : We agree that comparison with another fine-tuning method would make the real-robot results easier to interpret. We have therefore added a DMPO-style baseline that introduces stochasticity into the MeanFlow policy through a learned noise network. All other experimental settings and hyperparameters were kept identical to those used for MRFT. For each task, both methods were fine-tuned for the same wall-clock duration and evaluated in three rounds of 20 trials, totaling 60 trials per method per task. Our work focuses on deriving and theoretically analyzing a stochastic sampling rule for RL fine-tuning of MeanFlow policies, and this controlled comparison assesses its practical effectiveness.
Cube-to-Saucer Placement

MethodRound 1Round 2Round 3Mean success rateOverall
Pretrained MeanFlow5/20 (25%)5/20 (25%)4/20 (20%)23.3% ± 2.9%14/60
MeanFlow + DMPO-style learned-noise baseline10/20 (50%)12/20 (60%)9/20 (45%)51.7% ± 7.6%31/60
MeanFlow + MRFT12/20 (60%)14/20 (70%)13/20 (65%)65.0% ± 5.0%39/60

Cube-into-Cup Placement

MethodRound 1Round 2Round 3Mean success rateOverall
Pretrained MeanFlow9/20 (45%)8/20 (40%)7/20 (35%)40.0% ± 5.0%24/60
MeanFlow + DMPO-style learned-noise baseline13/20 (65%)13/20 (65%)14/20 (70%)66.7% ± 2.9%40/60
MeanFlow + MRFT14/20 (70%)17/20 (85%)16/20 (80%)78.3% ± 7.6%47/60

[W2.4] Real-robot experiments: the final performance after fine-tuning is rather limited (around 60%-70% success rate).
We agree that the reported success rates leave room for improvement. For each task, MRFT and the DMPO-style baseline were initialized from the same pretrained MeanFlow policy and fine-tuned for the same fixed wall-clock duration. Training ended when the allocated time was reached, rather than according to a convergence criterion. The reported rates therefore characterize performance within this training budget. Prior real-robot RL studies, including SHaRe-RL$^{[2]}$ and Event2Reward$^{[3]}$, similarly evaluate adaptation after training for fixed wall-clock durations.
Our contribution focuses on the derivation and analysis of a stochastic sampling rule for MeanFlow reinforcement fine-tuning, with the hardware experiments assessing whether it enables policy improvement through real-world interaction. Within the allocated budget, MRFT improved success over the pretrained policies by 35 and 25 percentage points on the saucer and cup tasks, respectively, in the original evaluation. To provide a clearer assessment, we have added the DMPO-style comparison under the same training budget and expanded evaluation to three rounds of 20 trials per method and task, reporting both individual rounds and the 60-trial results.


[1] π\pi-RL: Online RL Fine-tuning for Flow-based Vision-Language-Action Models, Arxiv 2025.
[2] SHaRe-RL: Structured, Interactive Reinforcement Learning for Contact-Rich Industrial Assembly Tasks, ICRA 2026.
[3] Event2Reward: Learning Robotic Rewards from Progress Relations, Arxiv 2026.

AuthorsOpenReview ↗

Response to Reviewer oxBK (Part 3)

[W3] Reproducibility: experiment details, how hyper-parameters are tuned for the proposed approach and the baselines are not specified.
Reply: We appreciate your feedback on reproducibility. Relative to DMPO, our method introduces only one additional hyperparameter, the noise level a, while retaining DMPO’s settings for the shared hyperparameters. We therefore kept the original description brief. But we agree that these settings and their selection should have been documented more explicitly. For the main experiments, we evaluated a ∈\in {0.2, 0.3, 0.4, 0.5} and selected the best-performing value. The search range was informed by the noise levels studied in π\pi-RL$^{[1]}$.
We used the official implementations and hyperparameter settings for all baselines. The only adjustment was the number of parallel environments for Robomimic, which we set to run on a single NVIDIA RTX 4090 GPU and kept identical across methods for each task. We have added the experimental setup and hyperparameter selection details to Appendix G of the revised manuscript.
[W4] Section 3 problem formulation: it seems that the paper is dealing with a finite-horizon MDP, then, the definition of the MDP tuple should include the horizon.
Reply: Thank you for pointing this out. We have revised the MDP tuple in Section 3 to explicitly include the horizon. You can find the updated definition in the Problem Formulation of the revised PDF. The passage now reads as follows:
We formulate robotic control as a finite-horizon Markov decision process M=(S,A,P0,Penv,R,γ,K+1)\mathcal{M}=(\mathcal{S},\mathcal{A},P_0,P_{\mathrm{env}},R,\gamma,K+1), where S\mathcal{S} and A\mathcal{A} denote the state and action spaces, P0P_0 is the initial-state distribution, PenvP_{\mathrm{env}} is the environment transition kernel, RR is the reward function, γ∈[0,1)\gamma\in[0,1) is the discount factor, and K+1K+1 is the episode horizon.


[1] π\pi-RL: Online RL Fine-tuning for Flow-based Vision-Language-Action Models, Arxiv 2025.

Reviewer_zu2GOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online reinforcement learning fine-tuning of few-step generative robotic policies based on MeanFlow. Given the initial noise, standard MeanFlow uses deterministic ODE updates and lacks stochastic transitions at individual sampling steps and tractable densities for these transitions, making it difficult to directly apply PPO-style optimization based on sampling-path likelihoods. Starting from a marginal-preserving Flow SDE, the authors derive MeanFlow-SDE through finite-interval integration and approximation, obtaining an explicit Gaussian transition with only one evaluation of the average velocity field per interval. They then factorize the joint density of the sampling path into a product of Gaussian transition densities, formulate a path-level clipped PPO objective, and introduce a Q-weighted actor-critic objective, forming the MRFT framework. The main contribution is the derivation of tractable stochastic transitions and path likelihoods from the underlying flow dynamics, providing a theoretical basis for online RL fine-tuning of MeanFlow. Experiments cover nine continuous-control tasks, LIBERO with a π0.5-style VLA policy, ablations on noise levels and sampling steps, and pick-and-place tasks on a Franka Research 3 robot. The results support effective fine-tuning with MRFT under a 2-NFE budget and its applicability to VLA policies and real robots.

Critical strengths
  1. Experimental validation covers few-step control, VLA policies, and real robots. The paper compares against DMPO and ReinFlow-R/S on nine continuous-control tasks, evaluates a π0.5-style VLA policy on LIBERO, and conducts online fine-tuning on a Franka Research 3 robot, alongside noise-level and NFE ablations. By reporting both task performance and online iteration time, the experiments provide good support for the effectiveness and applicability of few-step RL fine-tuning.

  2. The method construction has a clear theoretical basis. Through finite-interval analysis of the Flow SDE, the paper connects velocity integrals along stochastic trajectories to the MeanFlow average velocity. This yields a sampling mechanism requiring only one network evaluation per interval while providing an explicit Gaussian transition density, offering a concrete implementation for policy optimization based on path likelihoods.

Critical weaknesses

The approximation error in the practical two-step sampling setting remains unclear. The O((t - r)^(3/2 - ε)) result in the paper describes how the approximation error behaves as the step size approaches zero. However, the experiments primarily use 2-NFE, completing sampling in only two steps. With such large steps, this result alone does not establish whether the approximation error is sufficiently small. In other words, the existing experiments show that the method improves task performance, but the main text does not further analyze how large the numerical discrepancy is between the constructed transitions and those of the original Flow SDE when sampling in two steps. I suggest providing additional analysis or numerical comparisons to clarify how accurate this approximation is at the step sizes used in practice.

Overall assessment: Weak acceptance. The method has clear theoretical motivation and broad experimental validation, and its core contribution is valuable. There is still room to improve the connection between the local approximation analysis and the practical large-interval setting, as well as the statistical reporting of the real-robot results. These limitations are generally addressable, so I lean toward acceptance, but do not consider the evidence sufficient for a clear acceptance recommendation.

Questions for authors
  1. The theoretical error bound is a local asymptotic result. Could the authors use existing analysis or a small-scale numerical experiment to clarify the applicability of this approximation in the actual 2-NFE setting? For example, endpoint statistics of the constructed transitions could be compared with those of a fine-step reference SDE at the same coarse time points.

  2. The real-robot experiments involve 20 trials per task, with success rates increasing from 25% to 60% and from 45% to 70%. Could the authors report confidence intervals or results from additional trials, and describe the distribution of the initial object-pose perturbations? This would affect my assessment of the reliability of the real-robot conclusions.

4 comments
AuthorsOpenReview ↗

Response to reviewer zu2G (Part 1)

Thank you for your insightful comments and efforts in reviewing our manuscript. We deeply appreciate your valuable suggestions. In the following, we respond to each of your comments point by point.
[W1&Q1] The approximation error in the practical two-step sampling setting remains unclear. The O((t - r)^(3/2 - ε)) result in the paper describes how the approximation error behaves as the step size approaches zero. However, the experiments primarily use 2-NFE, completing sampling in only two steps. With such large steps, this result alone does not establish whether the approximation error is sufficiently small. In other words, the existing experiments show that the method improves task performance, but the main text does not further analyze how large the numerical discrepancy is between the constructed transitions and those of the original Flow SDE when sampling in two steps. I suggest providing additional analysis or numerical comparisons to clarify how accurate this approximation is at the step sizes used in practice. The theoretical error bound is a local asymptotic result. Could the authors use existing analysis or a small-scale numerical experiment to clarify the applicability of this approximation in the actual 2-NFE setting? For example, endpoint statistics of the constructed transitions could be compared with those of a fine-step reference SDE at the same coarse time points.
Reply: We agree with the reviewer that the additional quantitative analysis is needed for a direct assessment of how closely two-step MeanFlow-SDE matches the endpoint of a fine-step Flow-SDE reference. Following your suggestion, we conducted the proposed endpoint-statistics comparison using a frozen Square checkpoint: 2-step MeanFlow-SDE (1→0.5→01\rightarrow0.5\rightarrow0) versus a 128-step Flow-SDE reference using uθ(x,t,t)u_\theta(x,t,t) as the instantaneous velocity. Both used noise_level = 0.3. We evaluated 32 observations × 4 fixed initial latents (sampled from a standard Gaussian distribution), with 1024 stochastic realizations per observation–latent pair for each method.
Within each condition, we split each sampler’s draws into two batches of 512, labeled A and B. For the two cross-batch pairings (MeanFlow-SDE batch A with reference batch B, and MeanFlow-SDE batch B with reference batch A), we computed RMS differences in means and coordinate-wise variances across the 28 normalized coordinates (four actions × seven dimensions). We averaged over the two pairings, then over latents and observations.
RMS differences in means and diagonal variances (95% CIs in brackets).

\hline \text{Endpoint} & \text{Means} & \text{Diagonal variances} \\\\ \hline t = 0.5 & 0.0452\\;[0.0422,\\,0.0488] & 0.08021\\;[0.07995,\\,0.08045] \\\\ t = 0 & 0.0563\\;[0.0508,\\,0.0627] & 0.01015\\;[0.00802,\\,0.01278] \\\\ \hline \end{array}$$ **Two-step MeanFlow-SDE shows close agreement with the 128-step Flow-SDE reference in the final endpoint moments ($t=0$), with RMS differences of 0.0563 in means and 0.01015 in diagonal variances across the 28 normalized action coordinates. After inverse normalization and controller scaling to obtain the final actions, the average 3D distance between the final mean XYZ commands of 2-step MeanFlow-SDE and 128-step Flow-SDE is only 4.22 mm (95% CI: 3.95–4.51 mm).** These measurements complement our local asymptotic analysis by directly quantifying discrepancies in endpoint moments at practical step sizes.
AuthorsOpenReview ↗

Response to reviewer zu2G (Part 2)

[Q2] The real-robot experiments involve 20 trials per task, with success rates increasing from 25% to 60% and from 45% to 70%. Could the authors report confidence intervals or results from additional trials, and describe the distribution of the initial object-pose perturbations?
[Q2.1] Could the authors report confidence intervals or results from additional trials?.
Reply: Yes. We initially evaluated each method over 20 real-robot trials per task, following prior works. DPPO$^{[1]}$ used this protocol for One-leg assembly, and ConRFT$^{[2]}$ used it across eight manipulation tasks. We appreciate the reviewer’s suggestion that confidence intervals or additional trials would give a more precise estimate of the gains from fine-tuning. We conducted two additional rounds of 20 trials, bringing the total to 60 trials per method per task. Tables 1 and 2 summarize the results across all three rounds. We promise that we will add these results to the final version.

Table 1. Cube-to-Saucer Placement

\hline \textbf{Method} & \textbf{Round 1} & \textbf{Round 2} & \textbf{Round 3} & \begin{array}{l}\textbf{Average} \\\\ \textbf{success rate}\end{array} & \textbf{Overall} \\\\ \hline \begin{array}{l}\text{Pretrained} \\\\ \text{MeanFlow}\end{array} & 5/20\\;(25\\%) & 5/20\\;(25\\%) & 4/20\\;(20\\%) & 23.3\\% \pm 2.9\\% & 14/60 \\\\ \hline \begin{array}{l}\text{MeanFlow +} \\\\ \text{MRFT}\end{array} & 12/20\\;(60\\%) & 14/20\\;(70\\%) & 13/20\\;(65\\%) & 65.0\\% \pm 5.0\\% & 39/60 \\\\ \hline \end{array}$$ **Table 2. Cube-into-Cup Placement** $$\begin{array}{|l|l|l|l|l|l|} \hline \textbf{Method} & \textbf{Round 1} & \textbf{Round 2} & \textbf{Round 3} & \begin{array}{l}\textbf{Average} \\\\ \textbf{success rate}\end{array} & \textbf{Overall} \\\\ \hline \begin{array}{l}\text{Pretrained} \\\\ \text{MeanFlow}\end{array} & 9/20\\;(45\\%) & 8/20\\;(40\\%) & 7/20\\;(35\\%) & 40.0\\% \pm 5.0\\% & 24/60 \\\\ \hline \begin{array}{l}\text{MeanFlow +} \\\\ \text{MRFT}\end{array} & 14/20\\;(70\\%) & 17/20\\;(85\\%) & 16/20\\;(80\\%) & 78.3\\% \pm 7.6\\% & 47/60 \\\\ \hline \end{array}$$ **[Q2.2] Could the authors describe the distribution of the initial object-pose perturbations?** **Reply**: Yes. For both tasks, we varied the cube’s starting pose relative to its nominal pose on the table and kept the receiving object fixed. The angular offset refers to the cube’s initial yaw about the vertical axis. Further details are provided in Table 3. For each trial, we manually set the cube’s initial position and yaw within the ranges specified in Table 3. **Table 3. Initial object-pose perturbations and task demands** $$\begin{array}{|l|l|l|l|} \hline \textbf{Task} & \textbf{Task type} & \textbf{Initial object-pose perturbations} & \textbf{Key challenges} \\\\ \hline \begin{array}{l} \text{Cube-to-} \\\\ \text{Saucer} \\\\ \text{Placement} \end{array} & \text{Pick-and-place} & \begin{array}{l} \text{Cube position varied within 2 cm} \\\\ \text{along each of x and y. Cube yaw} \\\\ \text{varied by up to 5° clockwise or} \\\\ \text{counterclockwise. The saucer} \\\\ \text{remained fixed.} \end{array} & \begin{array}{l} \text{Grasping a small cube from a variable} \\\\ \text{starting position, keeping it stable} \\\\ \text{during transport, and placing it} \\\\ \text{accurately on the saucer.} \end{array} \\\\ \hline \begin{array}{l} \text{Cube-into-} \\\\ \text{Cup} \\\\ \text{Placement} \end{array} & \begin{array}{l} \text{Pick-and-place} \\\\ \text{with large-angle} \\\\ \text{wrist reorientation} \end{array} & \begin{array}{l} \text{Cube position varied within 1 cm} \\\\ \text{along each of x and y. Cube yaw} \\\\ \text{varied by up to 5° clockwise or} \\\\ \text{counterclockwise. The cup} \\\\ \text{remained fixed.} \end{array} & \begin{array}{l} \text{Executing a large inward wrist} \\\\ \text{rotation, grasping the cube with a} \\\\ \text{partially occluded wrist-camera view,} \\\\ \text{and aligning the cube with the cup} \\\\ \text{opening before release.} \end{array} \\\\ \hline \end{array}$$ --- [1] Diffusion policy policy optimization, ICLR 2025. [2] ConRFT: A Reinforced Fine-tuning Method for VLA Models via Consistency Policy, RSS 2025.

Thank you for your thoughtful and detailed responses, as well as the substantial effort devoted to the additional experiments. I particularly appreciate that the authors directly addressed both of my questions with new empirical evidence. These additions significantly strengthen the paper and provide a much clearer understanding of the proposed method.

  1. Approximation accuracy in the 2-NFE setting.

I appreciate the additional comparison between two-step MeanFlow-SDE and the 128-step Flow-SDE reference, which closely follows the numerical experiment suggested in my original review. The reported differences in endpoint means and variances, together with the 4.22 mm difference between the final mean XYZ commands, provide valuable empirical evidence for the accuracy of the approximation at practical step sizes.

Although the evaluation focuses on one task, checkpoint, and noise level, and matching the first two marginal moments does not fully characterize the underlying distributions, I believe the new experiment provides meaningful support for the practical validity of the proposed approximation. My original concern regarding the 2-NFE setting has been substantially addressed.

  1. Reliability of the real-robot experiments.

I also appreciate the authors' effort in expanding the real-robot evaluation from 20 to 60 trials per method per task. The consistent improvements across three rounds make the experimental conclusions considerably more convincing. The additional information about initial object-pose perturbations further improves the clarity and reproducibility of the experimental setup.

The new experimental results are already clearly presented and provide strong support for the effectiveness of the proposed fine-tuning approach. Two small clarifications could make the presentation even more informative:

  • Clarifying whether the ± values in Tables 1 and 2 represent standard deviations across rounds or another measure of variability.
  • Briefly explaining how the initial object poses were selected within the reported perturbation ranges, given that they were manually configured.

These are minor suggestions for improving the presentation rather than substantive concerns about the experimental conclusions.

Overall assessment.

I am pleased to see that the authors have carefully considered the review comments and provided substantial new evidence addressing both questions. The additional experiments significantly improve the connection between the theoretical analysis and practical few-step sampling, while also strengthening the reliability of the real-robot results.

I consider my main concerns satisfactorily addressed and maintain my original Weak Acceptance (3) recommendation, with increased confidence in the paper's contributions. I encourage the authors to incorporate these valuable additions into the revised manuscript.

Dear Reviewer zu2G,

Thank you very much for your thoughtful follow-up and for the time and care you have devoted to reviewing our work. We sincerely appreciate your acknowledgment that the additional experiments have satisfactorily addressed your main concerns.

Thank you also for your helpful suggestions. We apologize for not making these details sufficiently clear in our previous response. The “±” values in Tables 1 and 2 denote the standard deviations of success rates across the three evaluation rounds. The initial object poses were predefined within the reported perturbation ranges. We used small markers in the environment, such as black tape matching the color of the work surface, to indicate these predefined poses and manually positioned the objects accordingly. We will explicitly include these clarifications, together with the additional experimental results, in the revised manuscript.

We are grateful for your expressed increase in confidence in the paper’s contributions. If the additional evidence has also strengthened your confidence in your assessment, might you feel comfortable considering an increase in your confidence score? We would, of course, fully respect your judgment.

Best regards,
The Authors

Comments on the paper

1 comment

Dear Reviewers

The author-reviewer discussion period will close soon. Please read and respond to authors if you have not done yet.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.