acceptodds
Under review as a conference paper at ICLR 2027

EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control

Abstract

We introduce EfficientTDMPC, a sample-efficient model-based reinforcement learning method for continuous control built on the TD-MPC family of algorithms. Central to this family is a planner that aims to find an action sequence that maximizes the estimated return. The return is estimated using a learned model and value networks, each of which can introduce error. EfficientTDMPC introduces three contributions that improve performance by aiming to reduce this error. First, we introduce a multi-horizon planning objective that evaluates the value at different rollout depths and averages them. Second, to our knowledge we are the first to train a value-equivalent dynamics ensemble. Our improved objective then averages over rollouts from multiple dynamics heads. Third, we add pessimistic reanalyze for tasks that can terminate early. Applying our contributions to a recent baseline (BMPC) yields EfficientTDMPC, which to our knowledge is the new state of the art in sample efficiency on HumanoidBench and the DeepMind Control Suite, reaching BMPC's final aggregated performance using 57% fewer environment steps.

Primary area:
reinforcement learning
Keywords:
Sample efficient, Reinforcement learning, DMC, HumanoidBench, Deepmind Control Suite, Model-based, SOTA

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FK4cOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper builds on BMPC to propose EfficientTDMPC that makes planning less prone to model errors. It averages return estimates across rollout depths and dynamics models, and adds an uncertainty penalty when refreshing policy targets. On HumanoidBench and DMC, the resulting method reaches BMPC’s final aggregate performance with 57% fewer environment steps, while keeping training time roughly the same.

Critical strengths
  1. The method is straightforward and easy to understand. Both multi-horizon and multi-dynamics aggregation make intuitive sense to me.

  2. More importantly, the results and ablations show that these simple changes work well in practice. This is what matters most to me here, especially after seeing many recent methods struggle to make clear gains over Mrs. Q.

Critical weaknesses
  1. I would like to see a deeper analysis of why multi-horizon and multi-dynamics aggregation work. The cross-scoring analysis is a useful start, but I still do not have a clear picture of what drives the gains from each. For example, does multi-horizon aggregation help mainly by reducing estimation errors, or by changing which action sequences the planner prefers? The empirical results are convincing, and I think a clearer understanding of the mechanisms would add an important missing piece.
Questions for authors

See Weaknesses.

2 comments

Thank you for your review, apologies for the delayed response.

I would like to see a deeper analysis of why multi-horizon and multi-dynamics aggregation work. […] The empirical results are convincing, and I […] understanding of the mechanisms would add an important missing piece.

The effect of the dynamics aggregation is simple, we have multiple estimates of the same quantity and we average them for a lower error estimate. The cross scoring analysis in Figure 11 shows that the dynamics aggregation reduces estimation error and our ablations in Figure 4 show a significant increase in sample efficiency. The horizon aggregation is indeed more complex because it is a truely different objective that is being estimated and optimized. The contribution has two properties as you mentioned, 1) It changes the true objective, and 2) that objective can be estimated with less error.

We hypothesize that the sample-efficiency gain comes from lower estimation error. Firstly, the measurements we added in Appendix C.2 and the table below show that the objective is estimated with lower error, so we expect this to have an effect. Secondly, without estimation errors there is no reason to expect our horizon aggregated objective to choose better action sequences than a terminal value bootstrap, since the terminal value bootstrap is per definition the value of the full sequence, which is what we aim to maximize. Our horizon aggregated objective is only useful because it can be estimated with less error, but it does represent how good an action sequence is as well as a terminal value bootstrap.

A deeper experimental analysis of the multi-horizon objective is outside the scope of this paper. The table below shows how horizon aggregation reduces value estimation error during planning. More details are in Appendix C.2.

ReacherQuadrupedDogHumanoid
Correlation over dynamics heads0.560.610.750.86
Correlation over depth0.050.060.310.27
1 head, aggregate0.76 ± 0.110.74 ± 0.111.60 ± 0.103.33 ± 0.14
1 head, terminal1.97 ± 0.311.47 ± 0.183.38 ± 0.178.10 ± 0.32
4 heads, aggregate0.68 ± 0.110.41 ± 0.080.94 ± 0.092.33 ± 0.13
4 heads, terminal1.09 ± 0.130.73 ± 0.122.11 ± 0.156.57 ± 0.31

We hope this clarifies your concern and we hope you will consider raising your score.

Thanks for the rebuttal and the clearification, I decide to keep my score.

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

This paper proposes improvements to the TD-MPC family of latent world model-based reinforcement learning frameworks. Specifically, the authors introduce a multi-horizon planning objective and a value-equivalent dynamics ensemble to reduce the variance of return estimates, aiming to improve sample efficiency. Building on the ensemble-based approach, the proposed framework also estimates uncertainty in the learned dynamics and incorporates a pessimistic reanalysis mechanism to reduce the influence of high-uncertainty samples in the replay buffer.

The proposed method is evaluated against TD-MPC variants and other reinforcement learning algorithms, including Dreamer and SAC. The experimental results demonstrate improved sample efficiency compared to the considered baselines.

Critical strengths
  1. Improved sample efficiency. The experimental results demonstrate improvements in sample efficiency compared to both TD-MPC variants and other reinforcement learning baselines.

  2. Well integration of established techniques. The paper effectively incorporates concepts previously explored in reinforcement learning and model predictive control (MPC), such as dynamics ensembles and multi-horizon planning objectives, into the TD-MPC framework. This integration provides a coherent approach to improving the robustness of model-based planning.

  3. Clarity of presentation and experimental evaluation. Overall, the manuscript is well organized and relatively easy to follow. The three main contributions are clearly presented, and the experiments in Sections 5.2 and 5.3 provide useful evidence supporting the effectiveness of the individual components.

Critical weaknesses
  1. Limited theoretical justification. The manuscript provides limited theoretical insight into why the proposed methods improve sample efficiency. In particular, the mathematical relationship between the proposed multi-horizon planning objective, the value-equivalent dynamics ensemble, and the reduction in estimation variance is not sufficiently established. Furthermore, it remains unclear how reducing estimation variance translates into improved sample efficiency within the TD-MPC framework. A more rigorous theoretical analysis, or at least a discussion grounded in existing theoretical results, would strengthen the motivation for the proposed approach.

  2. Insufficient analysis of the underlying mechanisms. As discussed in the manuscript, multi-horizon planning and dynamics ensembles are intended to reduce estimation variance. However, the experimental evaluation primarily reports AUC and normalized returns without directly examining the estimation variance or bias of the proposed method. Additional analyses of estimation variance, estimation bias, and potential value overestimation would help clarify the relationship between the proposed mechanisms and the observed performance improvements. In particular, it is currently unclear whether the improved sample efficiency can be attributed to mitigating value overestimation, reducing estimation variance, or other factors introduced by the proposed framework. A more targeted empirical analysis would help substantiate the claimed benefits of the individual components.

  3. Lack of inference-time analysis. The proposed approach employs multiple dynamics models and additional value function evaluations, which may increase computational overhead during planning. However, the manuscript does not provide a detailed analysis of inference-time computational costs. Although wall-clock training time is reported under different reanalysis configurations, these results do not directly characterize the computational overhead introduced during online planning. A comparison of planning latency and computational complexity against the TD-MPC baselines would provide a more comprehensive assessment of the practical trade-offs associated with the proposed method.

Questions for authors
  1. Value overestimation versus estimation variance. The Introduction identifies value overestimation caused by inaccuracies in the learned dynamics and value models as a key challenge. However, the multi-horizon planning objective and value-equivalent dynamics ensemble appear to primarily address estimation variance. Could the authors clarify the relationship between variance reduction and the value overestimation problem highlighted in the Introduction?

  2. Contribution of the planning horizon. Table 2 shows that EfficientTD-MPC uses a longer planning horizon than BMPC (6 versus 3). Since longer planning horizons improve performance, it is unclear whether the reported gains arise from the proposed components or the increased horizon. Could the authors provide an ablation study without "practical improvement" to isolate the contributions of the multi-horizon objective and dynamics ensemble?

  3. Interpretation from an MPC perspective. In classical MPC, a terminal cost approximates the cost-to-go beyond the finite planning horizon. Similarly, TD-MPC uses a learned value function to estimate long-term returns. However, the proposed multi-horizon objective evaluates the Q-value function at multiple depths rather than only at the terminal state. Could the authors provide theoretical intuition for this design, particularly its relationship to conventional MPC terminal-cost formulations or multi-step Bellman targets?

  4. Minor presentation and notation issues.

    • Figure 2: The illustration appears to suggest that the reward is predicted by the value head. Clarifying the distinction between reward prediction and value estimation would improve readability.

    • Page 3: The notation (Xθ,i)(X_{\theta,i}) is introduced but does not appear to be clearly defined or consistently used elsewhere in the manuscript. It seems to represent a component of the latent dynamics model, but its distinction from (fθ,i)(f_{\theta,i}) is unclear. Could the authors clarify the meaning of these two terms?

    • Section 4.1: The notation (Q^aggregate,H)(\hat{Q}_{\mathrm{aggregate}, H}) or (Q^h)(\hat{Q}_{h}) appears potentially inconsistent with the surrounding definitions. Should the intended notation be (Q^hπ)( \hat{Q}_{h}^{\pi}) ? Clarifying this notation would improve readability and mathematical consistency.

5 comments

Thank you for your review, apologies for the delayed response.

high-uncertainty replay samples

We do not reduce the influence of high-uncertainty samples in the replay buffer. We penalize actions whose value is highly uncertain during planning. Those are different operations. Additionally, we compare to DreamerV3 not to Dreamer as the paper indeed states, we fixed that now.

limited theoretical justification

Appendix C.2 gives the explanation and the table. Averaging separate estimates whose errors are not fully correlated has lower error than any one estimate. The planner then keeps sequences whose true value is high, rather than sequences whose error is high, which raises sample efficiency.

estimation variance and bias

Table 3 in Appendix C.2 is repeated below. On checkpoints of the trained agent we plan from 512 states. The errors are not fully correlated across depths or dynamics heads, so averaging reduces the error. The estimation error of the selected sequences is lower with our contributions. That is lower variance and less overestimation, because maximization over high-variance estimates keeps the ones with large positive error. We hypothesize that this causes the sample-efficiency gains.

ReacherQuadrupedDogHumanoid
Correlation over dynamics heads0.560.610.750.86
Correlation over depth0.050.060.310.27
1 head, aggregate0.76 ± 0.110.74 ± 0.111.60 ± 0.103.33 ± 0.14
1 head, terminal1.97 ± 0.311.47 ± 0.183.38 ± 0.178.10 ± 0.32
4 heads, aggregate0.68 ± 0.110.41 ± 0.080.94 ± 0.092.33 ± 0.13
4 heads, terminal1.09 ± 0.130.73 ± 0.122.11 ± 0.156.57 ± 0.31

inference-time analysis

Our contributions do increase the compute cost of planning while acting in the environment, but this does not need to affect evaluation/inference latency. A core finding of BMPC (Wang et al., 2025) was that by distilling the planner into the policy network, the policy network performs almost equally well as planning. For this reason, for any compute sensitive application, the policy network should be used during evaluation, making the planning latency itself a much less relevant metric. We added this limitation to the conclusion and limitation section.

overestimation versus variance

Value overestimation and estimation variance are two sides of the same coin and we reduce both of them. Value overestimation/overestimation bias occurs when a maximisation operation, such as planning or policy improvement, is done over estimates with high variance. The maximisation focusses on estimates that happen to have high positive error, leading to overestimation bias. Our method reduces the estimation error, as we now show in Appendix C.2.

horizon of 6 versus 3

The contribution ladder in Figure 4 separates the longer acting horizon from the two new objectives. Adding horizon aggregation raises AUC by 11% on top of the practical changes. The dynamics ensemble raises it by a further 20%.

value at several depths, not only the terminal

Both methods aim to find a "good" action sequence and execute the first action. In classical MPC, a good action sequence is one that has the highest expected return when looking the furthest ahead, aka only bootstrapping from the terminal state. This is optimal when dynamics and terminal value estimates have no errors. We consider: what if there are estimation errors? In this case, we would want to use all the information we can to determine if an action sequence is good to take. The additional information that we use is: what is the value of just executing a part of the full action sequence? We prefer action sequences that have high estimated value even if only a part of it is executed. They are more robust to estimation errors. Consider two action sequences "A" and "B", both with equal estimated value. For sequence "A", executing only the first 5 actions is estimated to have very low value, while for sequence "B", executing only the first 5 actions is estimated to have high value as well. Then, clearly, in the presence of estimation errors we would prefer sequence "B". This is what our contribution adds to MPC, and our ablations in Figure 4 show that it can significantly improve sample efficiency. Multi-step bellman targets indeed also average over multiple depths, but they generally roll out policy actions, so each depth is an estimate of the same quantity, so averaging is a trivial variance reduction choice. Our method instead truely defines a different quantity to estimate.

Figure 2 and undefined notation

Thank you. We removed the unused symbols Xθ,iX_{\theta,i} and XθX_\theta. Q^h\hat{Q}_h is now defined as Q^hπ\hat{Q}^\pi_h, the return estimate defined in Section 2 at horizon hh. We will redraw Figure 2 so the reward prediction is distincter from the value heads.

We hope we clarified your concerns and hope you will consider raising your score.

Reviewer_fohmOpenReview ↗

Official comments by Reviewer fohm

Thank you for your response. Please see our comments below.

  • While App C.2 shows a reduction in value estimation error, we don't believe this necessarily implies a reduction in value estimation variance or overestimation bias in general.

  • The authors do not provide the requested theoretical justification for how the proposed methods reduce estimation variance and improve sample efficiency, instead relying on intuitive explanations and empirical results.

  • We agree that policy distillation may mitigate inference-time latency. However, our original concern was the inference latency of the proposed method itself, rather than potential approaches to mitigating this overhead. Moreover, the requested inference-time analysis has not been provided.

We believe that our main concerns remain unresolved, and therefore, we retain our original score.

Thank you for the follow-up.

While App C.2 shows a reduction in value estimation error, we don't believe this necessarily implies a reduction in value estimation variance or overestimation bias in general.

For any estimator Q^\hat{Q} of a target QQ,

E[(Q^−Q)2]⏟MSE=Var(Q^)⏟variance+(E[Q^]−Q)2⏟bias2.\underbrace{\mathbb{E}[(\hat{Q}-Q)^2]}_{\mathrm{MSE}} = \underbrace{\mathrm{Var}(\hat{Q})}_{\mathrm{variance}} + \underbrace{(\mathbb{E}[\hat{Q}]-Q)^2}_{\mathrm{bias}^{2}}.

The mean squared error is the variance plus the squared bias. The error therefore cannot fall unless the variance or the bias falls.

The authors do not provide the requested theoretical justification for how the proposed methods reduce estimation variance and improve sample efficiency, instead relying on intuitive explanations and empirical results.

Averaging errors that are not perfectly correlated reduces the variance part of the error. This follows from the variance formula, not from an empirical fit. For NN unbiased estimators of QQ, each with error variance σ2\sigma^{2} and pairwise correlation ρ\rho, the average is still unbiased and has variance

σ21+(N−1)ρN.\sigma^{2}\frac{1+(N-1)\rho}{N}.

This is strictly below σ2\sigma^{2} whenever ρ<1\rho < 1. When the errors are uncorrelated, ρ=0\rho=0 and the variance is σ2/N\sigma^{2}/N. Table 3 in Appendix C.2 shows ρ<1\rho < 1 across depths and across dynamics heads, so averaging over both dimensions reduces the estimator variance.

That lower variance improves the action the planner chooses. Take a simplified scenario with two actions with Q(a)>Q(b)Q(a) > Q(b), and estimates Q^(a)=Q(a)+ε(a)\hat{Q}(a)=Q(a)+\varepsilon(a) and Q^(b)=Q(b)+ε(b)\hat{Q}(b)=Q(b)+\varepsilon(b). Then the planner picks the incorrect (lower value) action if

ε(b)−ε(a)>Q(a)−Q(b).\varepsilon(b) - \varepsilon(a) > Q(a) - Q(b).

For independent Gaussian errors, ε(b)−ε(a)∼N(0,2σ2)\varepsilon(b)-\varepsilon(a)\sim\mathcal{N}(0,2\sigma^{2}), so

P(Incorrectly pick b)=Φ(−Q(a)−Q(b)σ2).P(\text{Incorrectly pick b}) = \Phi\left(-\frac{Q(a)-Q(b)}{\sigma\sqrt{2}}\right).

where Φ\Phi is the normal gaussian CDF.

This probability decreases with smaller estimator variance σ\sigma. A smaller estimation variance, as caused by our estimators, therefore in theory reduces the chance of the planner picking a lower value action. The picking of higher true value actions then per definition leads to the agent earning more reward throughout the entire learning process, leading to higher sample efficiency. Our empirical results back this theory up.

We agree that policy distillation may mitigate inference-time latency. However, our original concern was the inference latency of the proposed method itself, rather than potential approaches to mitigating this overhead. Moreover, the requested inference-time analysis has not been provided.

Policy distillation is in fact part of the baseline and our method, so it is not a potential mitigation, rather a reality of the method that does mitigate inference-time latency. We therefore do not do an additional latency analysis. We do mention an effect on inference time latency in the new limitations section as requested.

We hope this theory sufficiently supports the empirical sample efficiency gains of our contributions. We believe our contributions are valuable to the community and we hope you will consider recommending the paper to be accepted.

Reviewer_ZHC8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EfficientTDMPC improves the sample efficiency of BMPC through three changes to model-based planning: averaging bootstrapped returns across planning depths during acting, jointly training an ensemble of dynamics heads with shared reward and value networks, and applying an uncertainty penalty during reanalyze on tasks with early termination. The evaluation covers 14 HumanoidBench tasks and 28 DeepMind Control Suite tasks. The paper reports reaching BMPC’s final normalized aggregate performance with 57% fewer environment steps. The main contribution is an empirically effective combination of planning objectives and training modifications.

Critical strengths
  1. Compared to a powerful baseline model, the method achieves some empirical improvements, which are well-supported. Figure 3 and Appendix B show that the method learns faster on several challenging control tasks, with particularly notable improvements on the HumanoidBench task.

  2. The paper discusses the computational implications of its changes. This enhances the comprehensiveness of the paper. And figure 5 reports tasks on which pessimism hurts performance. This makes the proposed early-termination heuristic and its limitations more informative than a uniformly positive presentation would be.

Critical weaknesses
  1. The experimental results presented in the main text are promising; however, after reviewing the complete set of experiments in the appendix—such as the HumanoidBench results—the proportion of tasks showing significant performance improvements remains too small, so the overall performance improvement is not substantial enough.

  2. Comparisons with other methods also involve incomplete task coverage for DRQ and a different embodiment for SimBa-v2; describing the latter as an upper bound is not guaranteed merely because the embodiment is considered simpler.

  3. Some of the experimental figures in the paper have issues. For example, in Figure 9, although the number of experimental steps is only 2×10⁵, the same sampling interval as in the previous experiments was used for evaluation. Visually, this gives the impression of high variance. It is recommended that when the number of experimental steps is reduced, the number of evaluation data points be appropriately increased to make the curve smoother.

Questions for authors
  1. Some of the baselines in the paper are missing experimental results for certain tasks. How were the average scores for these baselines calculated? How was the impact of the missing tasks addressed? Is the average score for EfficientTDMPC calculated based on all 14 tasks? Were the missing results simply ignored, treated as zero, or handled in some other way?

  2. How are true terminal states, time-limit truncations, and termination during imagined rollouts handled in Eqs. (7) and (14)? No termination mask or termination model appears in these equations, although early termination determines when pessimism is enabled. Please specify the implementation and clarify how the task-selection rule and beta = 10 were chosen relative to the final evaluation.

  3. The paper claims that it “uses 57% fewer environmental interaction steps than BMPC.” Is this result robust, or could it be significantly influenced by the method used to calculate the composite score and by random variations in the experiments? The paper assigns equal weights to the “Easy” and “Hard” groups in DMC, but these groups consist of 21 and 7 tasks, respectively. Consequently, a single “Hard” task has three times the impact on the DMC composite score as an “Easy” task. If all tasks were weighted equally, would EfficientTDMPC’s advantage still be significant?

  4. The paper states that the training times for the two methods are similar. Under what conditions was this conclusion reached? How long does it actually take to select a single action? The fact that the total training times are similar does not mean that each action planning step is equally fast. EfficientTDMPC increases the number of dynamic models and the planning horizon in the acting phase, while reducing the computational load of the reanalyze phase. These two effects may offset each other, resulting in similar total training times, but individual action selection may still become slower. Therefore, how many milliseconds does it take, on average, to compute and output an action given the current observation? This information is very helpful for determining whether the method can be used for real-time control.

1 comment

Thank you for your review, apologies for the delayed response.

The experimental results presented in the main text are promising; however, after reviewing the complete set of experiments in the appendix—such as the HumanoidBench results—the propor […] significant performance improvements remains too small, so the overall performance improvement is not substantial enough.

Figure 1 in the attached pdf is an overview of the percent AUC improvement over BMPC (Wang et al., 2025) on every task, and on HumanoidBench, Hard DMC, Easy DMC, and the three suites together. The majority of tasks improve significantly due to our contributions, which yields the aggregate 2x sample efficiency improvement over the baseline.

Comparisons with other methods also involve incomplete task coverage for DRQ and a different embodiment for SimBa-v2; describing the latter as an upper bound is not guaranteed merely be […] embodiment is considered simpler.

For the missing DR.Q (Lyu et al., 2026) tasks we simply drop the tasks from the average. Since we normalize each task before averaging this does not change the expected value. We agree it is not fully valid so we mark it. We clarified this decision in the paper. For SimbaV2 (Lee et al., 2025), we removed the upper bound wording, and rather explained the with/without hands discrepancy in Appendix A.2.

Some of the experimental figures in the paper have issues. […] It is recommended that when the number of experimental steps is reduced, the number of evaluation data points be appropr […] make the curve smoother.

We choose to show 100k evaluation intervals because that is the data that is available for the baselines. Re-evaluating all baselines would be computationally expensive. The confidence intervals should give the reader all required information on the variance of the results.

Some of the baselines in the paper are missing experimental results for certain tasks. […] Were the missing results simply ignored, treated as zero, or handled in some other way?

We take the normalized average and skip tasks for which there are no results. Appendix A.2 makes this more clear.

How are true terminal states, time-limit truncations, and termination during imagined rollouts handled in Eqs. (7) and (14)? […] Please specify the implementation and clarify how the task-selection rule and beta = 10 were chosen relative to the final evaluation.

BMPC does not use a termination head, we do not change this. When a task can terminate early, we change the planning objective during reanalyze to include a pessimism term that subtracts 10 times the standard deviation over ensemble members. Time limit truncations of the episode, if that is what you mean, are handled in the standard way, by only selecting action sequences from the replay buffer that do not contain a truncation.

The paper claims that it "uses 57% fewer environmental interaction steps than BMPC." […] If all tasks were weighted equally, would EfficientTDMPC's advantage still be significant?

Yes the advantage remains under every weighting we checked. Equal weight on every task still reaches BMPC's final return in 47.5% fewer steps, and the AUC improvement is 44.4%. Equal weight on DMC and HumanoidBench reaches it in 55% fewer steps, and the AUC improvement is 53.6%. We choose to weigh each difficulty equally because it seems unfair to weigh easier tasks heavier simply because the benchmark has more of them.

The paper states that the training times for the two methods are similar. […] This information is very helpful for determining whether the method can be used for real-time control.

We trained 3 seeds of BMPC and EfficientTDMPC on the same task and reported the mean and standard deviation of the time it took to train 200k environment steps on an A100 for the hyper parameters reported in Table 2. Both methods require about the same wallclock per training step. The TDMPC family planner takes about 25ms on an A100 GPU to select an action. Our method increases the planning wallclock but this does not need to affect real-time control during evaluation or training. Regarding evaluation, a core finding of the BMPC paper was that by distilling the planner into the policy network, the policy network performs almost equally well as planning. For this reason, for real-time control the policy network should be used in either method. Regarding training, both BMPC and our method need to do live planning to act, reanalyze planning and training steps every environment step. EfficientTDMPC chooses to do more expensive live planning to act, whereas BMPC does more expensive reanalyze planning, but both methods require about the same wallclock per training step, and therefore have the same latency.

We hope we clarified your concerns and hope you consider increasing your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.