pith. sign in

arxiv: 2605.16692 · v2 · pith:BYCGRO4Ynew · submitted 2026-05-15 · 💻 cs.LG · cs.AI· cs.RO

EfficientTDMPC: Improved MPC Objectives for Sample-Efficient Continuous Control

Thomas Evers , Cristian Meo , Wendelin Bohmer , Justin Dauwels , Yaniv Oren This is my paper

Pith reviewed 2026-05-20 18:55 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.RO
keywords sample-efficient reinforcement learningmodel predictive controlcontinuous controlensemble methodsuncertainty-aware planningmodel-based RLTD-MPC
0
0 comments X

The pith

By averaging return estimates over model ensembles and rollout depths while penalizing uncertainty, EfficientTDMPC reaches leading sample efficiency on hard continuous control tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces EfficientTDMPC as a refinement of TD-MPC algorithms for model-based reinforcement learning in continuous control. It reduces errors in the planner's return estimates by averaging across an ensemble of learned dynamics models and across multiple rollout depths. An optional uncertainty penalty is added to discourage selection of actions whose returns are poorly estimated. Practical adjustments keep replay data fresher and lower unnecessary computation, which in turn lets the algorithm exploit higher update-to-data ratios. These changes together produce state-of-the-art sample efficiency on HumanoidBench-Hard and hard DMC benchmarks in the low-data regime while matching existing best results on easier DMC tasks.

Core claim

EfficientTDMPC improves the return estimation inside the model predictive controller by averaging the outputs of an ensemble of dynamics models over both the different models and different rollout depths, and by optionally subtracting an uncertainty term from the objective. Combined with updates that keep the replay buffer fresher and cut unnecessary computation, the algorithm can safely increase its update-to-data ratio and thereby learn more quickly from limited interaction data, reaching state-of-the-art sample efficiency on HumanoidBench-Hard and hard DeepMind Control tasks.

What carries the argument

Ensemble averaging of return estimates across multiple dynamics models and rollout depths, combined with an optional uncertainty penalty inside the planner objective.

If this is right

  • Higher update-to-data ratios become usable without causing instability or overfitting.
  • The planner selects actions on the basis of more reliable return estimates.
  • Sample-efficiency gains appear across both hard and easy continuous-control benchmarks.
  • Practical buffer and compute tweaks reduce the overhead of running model-based planning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same averaging and penalty ideas could be applied to other learned-model planners that currently rely on single-model rollouts.
  • In physical robotics the uncertainty penalty might lower the chance of executing actions that exploit model errors.
  • Scaling the ensemble size further could produce additional gains if the extra compute remains affordable.

Load-bearing premise

The ensemble averaging and uncertainty penalty genuinely reduce return estimation error rather than introducing compensating biases that only appear helpful on the tested benchmarks.

What would settle it

An ablation experiment on HumanoidBench-Hard in which removing either the ensemble averaging or the uncertainty penalty produces no measurable drop in sample efficiency would show that the claimed error reduction is not the operative mechanism.

Figures

Figures reproduced from arXiv: 2605.16692 by Cristian Meo, Justin Dauwels, Thomas Evers, Wendelin Bohmer, Yaniv Oren.

Figure 1
Figure 1. Figure 1: Mean normalized sample efficiency across DMC Easy, Hard and HumanoidBench￾Hard. Bars show mean area under the normalized aggregated learning curves(AUC) of each bench￾mark; error bars show 95% CIs over 35 tasks, 3 seeds each. Preprint. arXiv:2605.16692v2 [cs.LG] 19 May 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Process of return estimation of EfficientTDMPC vs BMPC (Right) BMPC rolls out a single dynamics head, predicts the reward earned at each depth and bootstraps with its value ensemble. (Left) EfficientTDMPC creates multiple rollouts using an ensemble of dynamics models. It then predicts the reward and bootstrapped value at each depth. The estimate is then averaged over the different rollouts and the return e… view at source ↗
Figure 3
Figure 3. Figure 3: gives the normalized aggregate learning curves of EfficientTDMPC compared to several strong baselines. In normalized aggregate learning curves, EfficientTDMPC reaches higher sample efficiency than the strongest compared baseline on HumanoidBench-Hard and Hard DMC, and remains competitive with the strongest baseline on Easy DMC. 0 2.5 5 7.5 10 0.0 0.5 1.0 1.5 2.0 Normalised Episode Return HumanoidBench 0 1.… view at source ↗
Figure 4
Figure 4. Figure 4: Component ablations. Left: the effect of the dynamics-ensemble size. Center: the effect of per step replay-buffer insertion. Right: The effect of the horizon aggregation. Shaded regions show 95% confidence intervals for mean normalized return across four ablation tasks with six seeds each. Reanalyze pessimism. We ablate the effect of pessimism during reanalyze by sweeping the pessimism coefficient β over {… view at source ↗
Figure 5
Figure 5. Figure 5: Reanalyze pessimism ablation. Evaluation reward for pessimism coefficients β ∈ {0, 1, 3, 10, 30} on h1hand-walk, dog-run, and reacher-hard, which are chosen to show a represen￾tative effect. Moderate pessimism improves h1hand-walk, has a mixed effect on reacher-hard, and degrades dog-run at larger coefficients. Combined Contributions and UTD scaling. As mentioned previously, the contributions of Effi￾cient… view at source ↗
Figure 6
Figure 6. Figure 6: UTD scaling and runtime. (a) Minutes to train for 200k environment steps on a single NVIDIA A100 GPU for BMPC, and for EfficientTDMPC at different UTD ratios; full comparison detail is in Appendix A.4. (b) UTD scaling on humanoid-walk. EfficientTDMPC benefits strongly from higher UTD, while BMPC improves more gradually under the same scaling; shaded regions show 95% confidence intervals using 5 seeds per m… view at source ↗
Figure 7
Figure 7. Figure 7: Planner ablations. (a) Cheap reanalyze: normalized aggregate planner performance showing the effect of cheaper reanalyze, comparing 512 reanalyze particles against 64 reanalyze particles across four HumanoidBench tasks, with 5 seeds per task for each condition, averaged after merging nearby evaluation checkpoints. Per task results are in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-task results: Humanoid-Bench (300k decision steps / 600k environment steps). Evaluation return on all 13 Humanoid-Bench tasks with all available baselines. The 7 tasks for which BOOM has data are used for the aggregate in the main paper, but we include all 13 tasks here for completeness. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Per-task results: Hard DMC (300k decision steps / 600k environment steps). Evaluation return on all 7 hard DMC tasks with all available baselines. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Per-task results: Easy DMC (100k decision steps). Evaluation return on all 21 easy DMC tasks with all available baselines. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-task component ablations. Learning curves for the isolated component ablations on quadruped-walk, reacher-hard, dog-stand, and humanoid-walk. Columns show the dynamics￾ensemble, replay-buffer-update, and horizon-aggregation ablations; shaded regions denote 95% confidence intervals over available seeds. B.5 Reduced reanalyze compute We experiment with the effect of reduced reanalyze compute on Efficien… view at source ↗
Figure 12
Figure 12. Figure 12: Reduced reanalyze compute does not degrade performance much. The per-task planner and policy learning curves show that training BMPC with 8x fewer reanalyze particles does not noticeably degrade performance on four HumanoidBench tasks with 61 action dimensions. The accompanying compact aggregate summarizes the normalized planner performance over the same four tasks. Shaded regions denote 95% confidence in… view at source ↗
Figure 13
Figure 13. Figure 13: HumanoidBench pessimism scope. Left: per-task evaluation return versus environment interactions on four HumanoidBench core tasks when pessimism is applied during reanalyze only versus during training-time planning, evaluation-time planning, and reanalyze. Right: the normalized aggregate over the same four tasks. Shaded regions denote 95% confidence intervals over available seeds. 21 [PITH_FULL_IMAGE:figu… view at source ↗
Figure 14
Figure 14. Figure 14: Ensemble averaging of return estimate reduces single-head planner exploitation. The plotted ∆R quantity estimates how much each planner is predicted to outperform the policy action from the same state. Bars show means over 512 replay states and error bars show standard errors over states. We find that the action sequences maximized under the single-head return estimate are given a very high return under t… view at source ↗
Figure 15
Figure 15. Figure 15: Latent trajectory gallery on reacher-hard at 50k (four states). Blue denotes actions from the optimistic planner (β = 0) and orange from the pessimistic planner (β = 10). The top row shows the latent trajectory in PCA space, both for each head’s predicted trajectory, the mean trajectory across heads, and the true environment. The bottom row shows the estimated return of each trajectory at each depth. 1.5 … view at source ↗
Figure 16
Figure 16. Figure 16: Latent trajectory gallery on quadruped-walk at 50k (four states). Same as [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Latent trajectory gallery on dog-stand at 50k (four states). Same as [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: DMControl tasks visualization. Images of all the embodiments we control in the DMControl tasks. The tasks include controlling them to run, walk, jump, balance, reach, and perform actions like swing-up and spin, covering a diverse range of continuous control scenarios. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: HumanoidBench locomotion suite visualization. Images of the Unitree robot we control in the HumanoidBench locomotion suite. The tasks include running, walking, crawling, balancing, sitting, reaching, and performing actions like walking on stairs or walking while avoiding collisions with poles, which cover a diverse range of robotic locomotion scenarios. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_19.png] view at source ↗
read the original 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 proposes to reduce this error in two ways. First, it introduces an ensemble of dynamics models and averages the return estimates across those models and across different rollout depths. Second, it adds the option to apply an uncertainty penalty to the planner objective, yielding a planner that avoids actions with uncertain return estimates. It then adds practical improvements which increase buffer data freshness and reduce compute. Lastly, we find that our contributions enable EfficientTDMPC to benefit more from a higher update-to-data (UTD) ratio, further improving sample efficiency. To the best of our knowledge, in the low data regime of each benchmark, EfficientTDMPC achieves state-of-the-art (SOTA) in terms of sample efficiency on HumanoidBench-Hard and DMC hard, while matching SOTA on DMC easy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces EfficientTDMPC, an extension of the TD-MPC family for sample-efficient model-based RL in continuous control. It proposes averaging return estimates over an ensemble of dynamics models and multiple rollout depths, an optional uncertainty penalty in the planner objective to avoid uncertain actions, practical improvements for buffer data freshness and reduced compute, and the ability to leverage higher update-to-data (UTD) ratios. The authors claim these changes yield state-of-the-art sample efficiency in the low-data regime on HumanoidBench-Hard and DMC hard tasks while matching SOTA on DMC easy.

Significance. If the empirical results are robust, the work offers a practical route to lower return-estimation error in MPC planning without requiring new model architectures. The ensemble averaging and uncertainty penalty are straightforward to implement and could be adopted by other model-based methods. No machine-checked proofs or parameter-free derivations are presented, but the focus on higher UTD ratios and data freshness provides concrete, falsifiable improvements that address known bottlenecks in sample-efficient RL.

major comments (2)
  1. [Abstract and §3] Abstract and §3: The central claim that ensemble averaging across models and rollout depths plus the uncertainty penalty produces lower-variance or lower-bias return estimates is load-bearing for the sample-efficiency results, yet the manuscript provides no direct quantitative evidence (e.g., MSE of estimated returns versus ground-truth rollouts or ablation curves isolating the averaging step). If model errors are correlated across depths, the averaging may not reduce error net of bias.
  2. [§5.2 and Table 3] §5.2 and Table 3: The reported SOTA claims on HumanoidBench-Hard and DMC hard in the low-data regime rest on the assumption that the uncertainty penalty weight and ensemble size do not require extensive per-benchmark tuning. The ablation results show performance sensitivity to these hyperparameters; without evidence that a single setting works across tasks, the headline sample-efficiency advantage is at risk of being undermined by hidden tuning costs.
minor comments (2)
  1. [Figure 2] Figure 2: Learning curves for the uncertainty-penalty ablation would benefit from shaded standard-error regions across all seeds to allow visual assessment of robustness.
  2. [§4.1] §4.1: The notation for the averaged return estimate (e.g., how depths are sampled and weighted) could be clarified with a short pseudocode block or explicit equation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and positive assessment of the practical contributions of EfficientTDMPC. We address each major comment below with clarifications and commit to revisions that strengthen the manuscript without overstating our results.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3: The central claim that ensemble averaging across models and rollout depths plus the uncertainty penalty produces lower-variance or lower-bias return estimates is load-bearing for the sample-efficiency results, yet the manuscript provides no direct quantitative evidence (e.g., MSE of estimated returns versus ground-truth rollouts or ablation curves isolating the averaging step). If model errors are correlated across depths, the averaging may not reduce error net of bias.

    Authors: We agree that direct quantitative evidence, such as MSE between estimated and ground-truth returns or isolated ablations of the averaging mechanism, would provide stronger support for the claim. The current manuscript offers indirect support via end-to-end sample-efficiency gains and ablations on ensemble size and the uncertainty penalty. To address the concern directly, including potential correlation of errors across rollout depths, we will add a new analysis section in the revision that reports return-estimation error metrics on a feasible subset of tasks using ground-truth rollouts. This addition will help verify net error reduction. revision: yes

  2. Referee: [§5.2 and Table 3] §5.2 and Table 3: The reported SOTA claims on HumanoidBench-Hard and DMC hard in the low-data regime rest on the assumption that the uncertainty penalty weight and ensemble size do not require extensive per-benchmark tuning. The ablation results show performance sensitivity to these hyperparameters; without evidence that a single setting works across tasks, the headline sample-efficiency advantage is at risk of being undermined by hidden tuning costs.

    Authors: We used a single fixed hyperparameter configuration—including the uncertainty penalty weight and ensemble size—across all tasks and benchmarks, with the exact values reported in the appendix. The ablations were performed to characterize sensitivity rather than to select per-task values. To clarify that the SOTA results do not rely on hidden per-benchmark tuning, we will revise the text in §5.2 and the appendix to explicitly state that hyperparameters were chosen once on representative tasks and transferred without further adjustment. This will better document the robustness of the reported gains. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical algorithmic improvements validated on external benchmarks

full rationale

The paper describes an extension of the TD-MPC family with ensemble averaging of return estimates across models and rollout depths, an optional uncertainty penalty, buffer freshness improvements, and higher UTD ratios. These are presented as practical algorithmic changes whose value is assessed through benchmark experiments on HumanoidBench and DMC. No equations, predictions, or first-principles claims are shown that reduce the reported performance gains to quantities defined by the method's own fitted parameters or prior self-citations. The central results rest on external empirical evaluation rather than any self-referential derivation, satisfying the self-contained criterion against benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 0 axioms · 0 invented entities

The approach rests on standard model-based RL assumptions plus two practical hyperparameters whose values are not derived from first principles.

free parameters (2)
  • ensemble size
    Number of dynamics models whose predictions are averaged; chosen to balance compute and error reduction.
  • uncertainty penalty weight
    Coefficient scaling the penalty term in the planner objective; tuned for performance on the target benchmarks.

pith-pipeline@v0.9.0 · 5742 in / 1101 out tokens · 48152 ms · 2026-05-20T18:55:22.249820+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 5 internal anchors

  1. [1]

    Machine Learning , volume =

    Bagging Predictors , author =. Machine Learning , volume =

    work page

  2. [2]

    International Conference on Machine Learning (ICML) , year =

    Temporal Difference Learning for Model Predictive Control , author =. International Conference on Machine Learning (ICML) , year =

    work page

  3. [3]

    Nicklas Hansen and Hao Su and Xiaolong Wang , booktitle =

    work page

  4. [4]

    2025 , eprint=

    TD-M(PC) ^2 : Improving Temporal Difference MPC Through Policy Constraint , author=. 2025 , eprint=

    work page 2025

  5. [5]

    The Thirteenth International Conference on Learning Representations , year =

    Bootstrapped Model Predictive Control , author =. The Thirteenth International Conference on Learning Representations , year =

    work page

  6. [6]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Bootstrap Off-policy with World Model , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

    work page

  7. [7]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Deep Reinforcement Learning in a Handful of Trials using Probabilistic Ensemble Trajectory Sampling , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  8. [8]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    When to Trust Your Model: Model-Based Policy Optimization , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  9. [9]

    Model-Based Value Estimation for Efficient Model-Free Reinforcement Learning

    Model-Based Value Estimation for Efficient Model-Free Reinforcement Learning , author =. arXiv preprint arXiv:1803.00101 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Sample-Efficient Reinforcement Learning with Stochastic Ensemble Value Expansion , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  11. [11]

    International Conference on Learning Representations (ICLR) , year =

    Dream to Control: Learning Behaviors by Latent Imagination , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  12. [12]

    Nature , volume =

    Mastering diverse control tasks through world models , author =. Nature , volume =. 2025 , month = apr, doi =

    work page 2025

  13. [13]

    International Conference on Learning Representations (ICLR) , year =

    High-Dimensional Continuous Control Using Generalized Advantage Estimation , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  14. [14]

    Randomized Ensembled Double

    Xinyue Chen and Che Wang and Zijian Zhou and Keith Ross , booktitle =. Randomized Ensembled Double

    work page

  15. [15]

    IEEE Transactions on Computational Intelligence and AI in Games , volume =

    A Survey of Monte Carlo Tree Search Methods , author =. IEEE Transactions on Computational Intelligence and AI in Games , volume =

    work page

  16. [16]

    2025 , eprint=

    Masked Generative Priors Improve World Models Sequence Modelling Capabilities , author=. 2025 , eprint=

    work page 2025

  17. [17]

    Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

    Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm , author =. arXiv preprint arXiv:1712.01815 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Mastering

    Julian Schrittwieser and Ioannis Antonoglou and Thomas Hubert and Karen Simonyan and Laurent Sifre and Simon Schmitt and Arthur Guez and Edward Lockhart and Demis Hassabis and Thore Graepel and Timothy Lillicrap and David Silver , journal =. Mastering

    work page

  19. [19]

    Rehg and Evangelos A

    Grady Williams and Paul Drews and Brian Goldfain and James M. Rehg and Evangelos A. Theodorou , journal =. Information Theoretic. 2018 , doi =

    work page 2018

  20. [20]

    Soft Actor-Critic Algorithms and Applications

    Soft Actor-Critic Algorithms and Applications , author =. arXiv preprint arXiv:1812.05905 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    International conference on machine learning , pages=

    Addressing function approximation error in actor-critic methods , author=. International conference on machine learning , pages=. 2018 , organization=

    work page 2018

  22. [22]

    Wurman and Jaegul Choo , booktitle =

    Hojoon Lee and Dongyoon Hwang and Donghu Kim and Hyunseung Kim and Jun Jet Tai and Kaushik Subramanian and Peter R. Wurman and Jaegul Choo , booktitle =

    work page

  23. [23]

    Hyperspher- ical normalization for scalable deep reinforcement learning.arXiv preprint arXiv:2502.15280,

    Hyperspherical Normalization for Scalable Deep Reinforcement Learning , author =. arXiv preprint arXiv:2502.15280 , year =

    work page arXiv

  24. [24]

    Yuval Tassa and Yotam Doron and Alistair Muldal and Tom Erez and Yazhe Li and Diego de Las Casas and David Budden and Abbas Abdolmaleki and Josh Merel and Andrew Lefrancq and Timothy Lillicrap and Martin Riedmiller , journal =

    work page

  25. [25]

    Machine Learning , volume =

    Learning to Predict by the Methods of Temporal Differences , author =. Machine Learning , volume =

    work page

  26. [26]

    Tianhe Yu and Garrett Thomas and Lantao Yu and Stefano Ermon and James Zou and Sergey Levine and Chelsea Finn and Tengyu Ma , booktitle =

    work page

  27. [27]

    International Conference on Learning Representations , year =

    Epistemic Monte Carlo Tree Search , author =. International Conference on Learning Representations , year =

    work page

  28. [28]

    arXiv preprint arXiv:2406.01423 , year =

    Value Improved Actor Critic Algorithms , author =. arXiv preprint arXiv:2406.01423 , year =

    work page arXiv

  29. [29]

    arXiv preprint arXiv:2511.14220 , year =

    Twice Sequential Monte Carlo for Tree Search , author =. arXiv preprint arXiv:2511.14220 , year =

    work page internal anchor Pith review arXiv

  30. [30]

    2026 , eprint =

    The Surprising Difficulty of Search in Model-Based Reinforcement Learning , author =. 2026 , eprint =

    work page 2026

  31. [31]

    SIGART Bulletin , volume =

    Dyna, an Integrated Architecture for Learning, Planning, and Reacting , author =. SIGART Bulletin , volume =

    work page

  32. [32]

    World Models

    World Models , author =. arXiv preprint arXiv:1803.10122 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Robotics: Science and Systems (RSS) , year =

    HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation , author =. Robotics: Science and Systems (RSS) , year =

    work page

  34. [34]

    2025 , eprint=

    Bigger, Regularized, Categorical: High-Capacity Value Functions are Efficient Multi-Task Learners , author=. 2025 , eprint=

    work page 2025

  35. [35]

    2017 , eprint=

    Proximal Policy Optimization Algorithms , author=. 2017 , eprint=

    work page 2017

  36. [36]

    2023 , eprint=

    Bridging Offline Reinforcement Learning and Imitation Learning: A Tale of Pessimism , author=. 2023 , eprint=

    work page 2023