Surrogate-Assisted Evolutionary Reinforcement Learning Based on Autoencoder and Hyperbolic Neural Network

Aimin Zhou; Bingdong Li; Hong Qian; Ke Tang; Mei Jiang; Peng Yang

REVIEW 5 major objections 7 minor 5 references

Surrogate-Assisted Evolutionary Reinforcement Learning Based on Autoencoder and Hyperbolic Neural Network

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AE-HNN-NCS combines a learned autoencoder and a hyperbolic classifier to pre-select policies, using about one-third of real environment evaluations while outperforming prior methods on Atari and MuJoCo.

desk verdict A useful surrogate-assisted ERL idea undermined by an unequal-budget comparison in the headline Atari table; the equal-budget ablations suggest the components help, but the broad performance claims don't hold as presented. read the letter →

arxiv 2505.19423 v2 pith:ZFCIXQUW submitted 2025-05-26 cs.LG cs.AI

Bingdong Li , Mei Jiang , Hong Qian , Ke Tang , Aimin Zhou , Peng Yang This is my paper

classification cs.LGcs.AI

keywords surrogate-assistedevolutionaryreinforcementlearningautoencoderhyperbolicneuralnetworkpolicyembeddingnegativelycorrelatedsearchAtariMuJoCosampleefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes AE-HNN-NCS, a surrogate-assisted evolutionary reinforcement learning method that tackles the high cost of evaluating deep-network policies. An autoencoder compresses million-parameter policies into a low-dimensional latent space, and a hyperbolic neural network classifies candidates as promising or unpromising so that only the most promising one is evaluated in the real environment each generation. On 10 Atari and 4 MuJoCo tasks, the method surpasses previous RL and evolutionary RL baselines while cutting real evaluations to roughly one-third, reducing wall-clock training time by 38% on average. The paper also argues that pre-selection in the reduced space works because the embedding preserves fitness rankings, not exact fitness values, and that reproduction must stay in the original high-dimensional space to avoid reconstruction errors.

What carries the argument

The load-bearing mechanism is the joint AE-HNN pre-selection loop. The autoencoder $\phi: \mathbb{R}^n \to \mathbb{R}^m$ with $n \gg m$ learns to compress policy weight vectors by minimizing reconstruction error; its encoder alone is used to project all sampled offspring. The hyperbolic neural network, using Poincaré ball geometry and Möbius operations, maps the latent codes into hyperbolic space and classifies each as promising ($f(x) \geq$ population average) or not; the classifier's continuous output probability provides a soft ranking. The candidate with the highest predicted probability is the single policy evaluated in the environment, then selected or rejected by the NCS update rule with the 1/5 success rule adjusting the sampling covariance. The autoencoder's decoder is deliberately unused, avoiding reconstruction error; the chosen index is mapped back by table lookup to the original high-dimensional policy.

What would settle it

Take a control task where the fitness landscape is not low-rank—say, a policy whose performance depends on noise in rarely visited dimensions—and measure Spearman's rank correlation between true fitness and AE-HNN predicted scores on a held-out pool of candidates. If the correlation is zero or negative while the method still outperforms NCS, the claimed mechanism is wrong; alternatively, if a fixed random projection with the same latent dimension achieves equal rank correlation and equal performance, the autoencoder's learned embedding is not the cause.

Watch

Extended reading notes

Core claim

The central claim is that high-dimensional evolutionary policy search can be made far more sample-efficient by a fully learnable two-stage surrogate: an autoencoder that maps policies into a low-dimensional latent code, and a hyperbolic neural network that ranks those codes by the probability of belonging to the good class. The HNN's soft classification score orders candidates, and only the top candidate is evaluated with the true simulator, while the parent population evolves in the original weight space using negatively correlated search. Experiments across 10 Atari games and 4 MuJoCo locomotion tasks show consistent improvement over gradient-based RL (A3C, PPO), evolutionary baselines (CES, NCS), and prior surrogate-assisted ERL (PE-SAERL), together with a measured reduction of real evaluations to about one-third. The paper further establishes, through rank-correlation experiments on six Atari games, that the method's validity rests on rank preservation: the low-dimensional embedding must keep the relative ordering of policy fitness close to the true ordering, which is easier than preserving exact fitness values and is achieved by the learned AE rather than a fixed random projection.

Load-bearing premise

The load-bearing premise is that the autoencoder's low-dimensional embedding preserves the fitness ranking of policies, so a classifier trained in that space ranks candidates the same way the true environment would.

Editorial extensions

If this is right

If the method holds up, evolutionary RL becomes practical for settings where environment rollouts dominate cost: the same policy quality with roughly one-third of the evaluations.
The rank-preservation analysis implies a design rule for surrogate-assisted ERL: pre-select in embedded space, reproduce in original space, and never decode latent codes to generate offspring.
Because the modules are learnable and pretrained without reward labels (random-agent interaction data suffices for the AE), the approach could transfer to new tasks where reward evaluation is the bottleneck.
The success on Montezuma's Revenge and Venture, sparse-reward Atari games, suggests the surrogate filter does not merely interpolate near previously seen good policies but actively guides exploration toward structured, hierarchical policy regions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper leaves implicit is replacing the binary promising/unpromising label with a calibrated regressor or an ordinal ranking head; the authors themselves note this as future work, and a regression variant would make the soft-ranking step more directly grounded in fitness.
The rank-preservation condition could be turned into an online diagnostic: monitor Spearman correlation between AE-embedding distances and true fitness each generation, and fall back to random projection or full evaluations when the correlation collapses.
Because the AE is trained on randomly generated policies, the embedding may become stale as the population converges; an online-updated or continual autoencoder could be a testable improvement.
The Atari results show performance closest to PPO on Enduro and best on Bowling and Freeway, hinting that the method favors tasks with simple dynamics and clear reward structure; a broader benchmark with varied environment stochasticity would test this boundary.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core idea is real: learn a low-dimensional policy embedding with an autoencoder and use a hyperbolic-network classifier to pre-select candidate policies before expensive rollouts. The equal-budget ablation against PE-SAERL—all at 0.1B frames—shows the AE and HNN each add something, and the rank-consistency study across three strategies is a useful conceptual contribution. That part of the paper is worth reading.

The main Atari comparison is compromised, though. Table I gives AE-HNN-NCS 0.1B environment frames while PLASTIC, STORM, PPO, and A3C get 40M, a 2.5x budget advantage, and Section IV-C states a different ratio (25M vs 10M). Either way, the proposed method receives more real-environment budget than the strongest baselines, so the reported margins cannot be attributed to the AE or HNN modules. Without equal-frame comparisons, the central 'significantly outperforms' claim does not hold.

The novelty claim is also overstated. The paper calls itself the first learnable policy embedding and surrogate for high-dimensional ERL, but it cites Wang et al. 2022, which uses a VAE-GP surrogate—that's a learnable embedding and regression model. The specific AE+HNN combination is new, but the 'first' claim is not.

Other soft spots: the 20 repeated tests appear to be repeated rollouts of one trained policy rather than independent training runs, so there are no error bars or significance tests. No code or data are provided. The MuJoCo table has no budget information, so I cannot tell whether those comparisons are fair.

The paper is honest about its scope—it is EA-based policy optimization, not hybrid ERL—and the conclusion flags limitations. But the evidence as presented does not support the headline claim.

Who this is for: researchers working on surrogate-assisted ERL or neuroevolution will find the ablation and rank-consistency analysis informative. The paper deserves a serious referee, but it needs major revision: equal-frame comparisons, independent seeds, significance testing, and a corrected novelty statement. I would not cite it in its current form.

Referee Report

5 major / 7 minor

Summary. AE-HNN-NCS is an evolutionary policy-search algorithm that uses an autoencoder to reduce the parameter vector of a policy network to a low-dimensional embedding and a hyperbolic neural network classifier to pre-select which sampled offspring is worth a real environment evaluation; NCS is the base optimizer. The paper reports experiments on 10 Atari games and 4 MuJoCo tasks, claiming that the method significantly outperforms existing RL and ERL baselines while using roughly one-third of the real evaluations, and that rank preservation in the embedded space explains the success. The paper also presents ablations separating the AE and HNN contributions and visualizations of policy trajectories.

Significance. Because evolutionary policy search on high-dimensional policies is expensive, a learnable embedding plus surrogate pre-selection would be practically valuable, and the empirical rank-consistency story addresses a real design question raised by PE-SAERL. The ablation in Table II and the explicit Algorithm 2 are useful elements. However, the headline claim of significant outperformance is not supported as presented: the main cross-method comparison uses unequal environment budgets, and the results lack variance information. The work therefore cannot currently justify the claimed significance, though the equal-budget in-family comparisons suggest the components merit further study.

major comments (5)

[Table I / Section IV-C] The central comparison is confounded by unequal environment budgets. Table I's Time row allocates 0.1B frames to AE-HNN-NCS, NCS, PE-FCPS-NCS, and CES, but only 40M frames to PPO, A3C, PLASTIC, and STORM; Section IV-C states the intended allocation as 25M and 10M, respectively. Under either reading, AE-HNN-NCS receives 2.5 times as many environment interactions as the strongest gradient-based baselines. Because the headline claim is that AE-HNN-NCS significantly outperforms prior RL and ERL methods, the reported margins on games such as Alien and SpaceInvaders cannot be attributed to the AE or HNN modules; a pure budget difference could produce similar improvements. The equal-budget comparisons in Table II are more informative, but they do not rescue the broad cross-method claim.
[Section V-A / Tables I-III] There are no error bars or significance tests. Section V-A says that a policy model was trained and 20 repeated tests were performed on that trained model, which reflects evaluation noise rather than variance across independent training runs. Tables I, II, and III report single average values with no standard deviations, confidence intervals, or significance tests, despite the abstract's claim of significant outperformance. Without multiple independent runs and appropriate statistical testing, the performance differences reported in the paper are not statistically supported.
[Section V-A / Figure 3] The key efficiency claim that real environment evaluations are reduced to nearly one-third and that wall-clock time is reduced by 38% is supported only by Figure 3. No table of evaluation counts, wall-clock times, or repeated-run statistics is provided. Since this is a central motivation for the method, the claim needs controlled quantitative evidence rather than a single qualitative figure.
[Section III-B / Section V-C] The rank-preservation assumption is validated only post hoc on the same games used in the main evaluation and method motivation. Section III-B explicitly relies on the assumption that the low-dimensional embedding preserves policy fitness ranks, and Section V-C tests this on six Atari games from the same benchmark. The additional assumption that the HNN's classification probability orders policies by true fitness, which underlies the soft-ranking selection, is not independently tested or proven. A more convincing treatment would evaluate rank preservation on held-out games or provide a concrete bound relating rank preservation to pre-selection quality.
[Section III-D / Algorithm 1] The pseudocode and the prose disagree on the pre-selection criterion. Algorithm 1 sorts the predicted labels and returns the first entry, while Section III-D states that the HNN's continuous probability score is used to establish a soft ranking over candidate policies. Sorting binary labels does not order candidate policies within the 'good' class by confidence, and the pseudocode does not show where the probability score enters the selection. This ambiguity is central to the surrogate module and makes the method as written not fully reproducible.

minor comments (7)

[Section V-F] The section title uses 'Mujoco' rather than the standard 'MuJoCo' spelling.
[Table I] The BeamRider Best value for PLASTIC, 8263.9, is far above its Average of 798.6 and above every other method's Best value; this appears to be a typo and should be checked.
[Algorithm 2] Line 10 of Algorithm 2 says 'as shown in Algorithm 2' but the referenced pre-selection routine is Algorithm 1.
[Section III-D / Algorithm 1] Algorithm 1 lists yi and fi as inputs but the pseudocode body refers to xi and labels computed from xi and fi; Equation (3) also applies f to yi,j even though f is defined for high-dimensional policies x. The notation should be aligned.
[Table II] The column header 'HNN- NCS' has a spurious space and is not explained; the ablation naming should be clarified.
[References / Appendix] Appendices A and B are referenced in Sections IV-C and V-F but are not included in the supplied text; the manuscript should either include them or remove the references.
[Section I / Section II] The text states that the method belongs to the first taxonomy category, EA-based policy optimization, while the title and abstract use the ERL label; these terminological choices should be harmonized.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the AE-HNN surrogate is trained on reconstruction loss and fitness-derived labels, while the headline results are benchmarked against external baselines; self-citations supply the base framework but are not load-bearing.

full rationale

No step in the claimed derivation reduces to its own input by construction. AE-HNN-NCS is an empirical pipeline: the autoencoder is trained by reconstruction loss (Eq. 2), the HNN surrogate is trained on binary labels defined by fitness relative to the batch average (Eqs. 3-4), and the final policy is scored by environment reward. The headline comparison in Table I is judged against external baselines (CES, PPO, A3C, NCS, PE-SAERL, PLASTIC, STORM) and the MuJoCo table against TD3, PPO, TERL, PROTECTED, and RE2-ERL, so the central outperformance claim is not a restatement of the training objective. The strongest candidate for a circular step is the Section V-C rank-consistency study: the low-dimensional ranking is produced by the HNN, which was trained on fitness-derived labels, so part of the measured Spearman/Kendall agreement reflects surrogate fit rather than an independent property of the AE embedding. However, this is a post-hoc sanity check of the 'when it works' condition, not a fitted parameter renamed as the headline result; the AE itself is trained without fitness information, so the consistency result is not forced by construction. Self-citations of PE-SAERL, NCS, and Yang et al. 2021 supply the base framework and motivation, but PE-SAERL and NCS also serve as empirical baselines, so the citations are not load-bearing. Finally, the unequal frame budgets in Table I and Section IV-C (0.1B vs 40M; 25M vs 10M) and the lack of independent-seed error bars are substantive fairness concerns, but they concern external validity rather than circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method rests on several unproven domain assumptions about policy space structure, including rank preservation under embedding, subset-feature sufficiency, and HNN confidence ordering. These are not derived from first principles, and only the rank assumption receives a post hoc empirical check.

free parameters (5)

Embedding dimension m
Dimension of the autoencoder latent space; chosen by hand and not specified in the main text, yet it controls how much policy structure is retained.
Number of candidate policies M per parent
Sampling budget for HNN pre-selection in Algorithm 2; value not reported.
HNN curvature c
Curvature of the Poincaré ball model used by the hyperbolic network (Section III.D); not specified.
NCS parameters (population size N, trade-off phi, covariance update)
Standard NCS hyperparameters that determine exploration-exploitation balance and are not given in the main text.
Autoencoder training budget and architecture
Depth, width, and training epochs for the AE are not reported, though they affect the quality of the embedding.

assumptions (5)

domain assumption The true reward of an RL policy is influenced only by a subset of policy features, so dimensionality reduction can retain the information that determines fitness.
Invoked in Section III.C to justify why an autoencoder can compress policy networks without losing performance-relevant signal.
ad hoc to paper There exists a policy embedding that approximately preserves the fitness ranking between the high-dimensional and low-dimensional spaces.
Stated in Section III.B as a requirement for pre-selection to work; the paper only verifies this empirically on a subset of games after the fact.
domain assumption The HNN's predicted probability of belonging to the promising class orders policies by their true fitness.
The surrogate uses soft ranking in Section III.D, so selection quality depends on this monotonicity assumption, which is not proven.
domain assumption The Poincaré ball geometry is better suited than Euclidean space for representing the hierarchical structure of high-dimensional policy spaces.
The motivation for choosing HNN over Euclidean classifiers in Section III.D; no quantitative evidence is provided.
domain assumption Environmental rollouts provide a reliable ground-truth fitness function f for each policy.
The evaluation function f is treated as exact in Algorithm 2; stochasticity in Atari evaluation is acknowledged but not modeled with error bars.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Surrogate-Assisted Evolutionary Reinforcement Learning Based on Autoencoder and Hyperbolic Neural Network." pith.science (2026). https://pith.science/paper/ZFCIXQUW

@misc{pith2026250519423,
  author       = {Pith},
  title        = {Pith review of: Surrogate-Assisted Evolutionary Reinforcement Learning Based on Autoencoder and Hyperbolic Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZFCIXQUW}},
  note         = {Machine review of arXiv:2505.19423}
}

read the original abstract

Evolutionary Reinforcement Learning (ERL), training the Reinforcement Learning (RL) policies with Evolutionary Algorithms (EAs), have demonstrated enhanced exploration capabilities and greater robustness than using traditional policy gradient. However, ERL suffers from the high computational costs and low search efficiency, as EAs require evaluating numerous candidate policies with expensive simulations, many of which are ineffective and do not contribute meaningfully to the training. One intuitive way to reduce the ineffective evaluations is to adopt the surrogates. Unfortunately, existing ERL policies are often modeled as deep neural networks (DNNs) and thus naturally represented as high-dimensional vectors containing millions of weights, which makes the building of effective surrogates for ERL policies extremely challenging. This paper proposes a novel surrogate-assisted ERL that integrates Autoencoders (AE) and Hyperbolic Neural Networks (HNN). Specifically, AE compresses high-dimensional policies into low-dimensional representations while extracting key features as the inputs for the surrogate. HNN, functioning as a classification-based surrogate model, can learn complex nonlinear relationships from sampled data and enable more accurate pre-selection of the sampled policies without real evaluations. The experiments on 10 Atari and 4 Mujoco games have verified that the proposed method outperforms previous approaches significantly. The search trajectories guided by AE and HNN are also visually demonstrated to be more effective, in terms of both exploration and convergence. This paper not only presents the first learnable policy embedding and surrogate-modeling modules for high-dimensional ERL policies, but also empirically reveals when and why they can be successful.

Figures

Figures reproduced from arXiv: 2505.19423 by the authors.

**Figure 1.** The proposed ERL framework integrates an Autoencoder and a Hyperbolic Neural Network for enhanced dimensionality reduction and policy pre-selection. The left part shows the traditional ERL workflow with EA and RL phases. Our algorithm innovates in the EA phase. We first use an Autoencoder to embed offspring policies to a lower-dimensional space. Then, the Hyperbolic Neural Network acts as a surrogate model, predicti… view at source ↗

**Figure 2.** The overview of autoencoder. justifiable for DNN-based policies. DNN models often have redundant weights from the convolutional layer to the fullyconnected layer, and a large number of neuron activations tend to approach zero. Even after removing these neurons, the model can still maintain the same expressive power, a phenomenon known as over-parameterization, with the corresponding technique called model pruning[… view at source ↗

**Figure 3.** Performance Analysis of ERL with NCS and AE-HNN [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: The ranking consistency of final performances among [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: The t-SNE visualization results of HNN-based policy pre-selection on three Atari games: Alien, Pong, and Freeway. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: The t-SNE visualization results of four algorithms (PLASTIC, STORM, AE-HNN-NCS, PE-HNN-NCS) on three Atari [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

5 extracted references · 2 canonical work pages

[2009]

Progress in Aerospace Sciences 45 (2009), 50–79

Recent advances in surrogate-based optimization. Progress in Aerospace Sciences 45 (2009), 50–79. [Francon et al.(2020)] Olivier Francon, Sergio Gonzalez, Babak Hodjat, El- liot Meyerson, Risto Miikkulainen, Xiaoyong Qiu, and Hojjat Shahrzad

work page 2009
[2016]

arXiv preprint arXiv:1606.01540

OpenAI Gym. arXiv preprint arXiv:1606.01540. [Cetin et al.(2022)] Edoardo Cetin, Benjamin Chamberlain, Michael Bron- stein, and Jonathan J. Hunt. 2022. Hyperbolic deep reinforcement learning. arXiv preprint arXiv:2210.01542. [Chrabaszcz et al.(2018)] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. 2018. Back to basics: Benchmarking canonical evoluti...

arXiv 2022
[2020]

In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO)

Effective reinforcement learning through evolutionary surrogate- assisted prescription. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO) . ACM, 814–822. [Fujimoto et al.(2018)] Scott Fujimoto, Herke Hoof, and David Meger. 2018. Addressing function approximation error in actor-critic methods. In International Conference on Mach...

arXiv 2018
[2024]

Advances in Neural Information Processing Systems 36 (2024)

Plastic: Improving input and label plasticity for sample efficient reinforcement learning. Advances in Neural Information Processing Systems 36 (2024). [Li et al.(2024)] Bingdong Li, Zixiang Di, Yanting Yang, Hong Qian, Peng Yang, Hao Hao, Ke Tang, and Aimin Zhou. 2024. It’s morphing time: Unleashing the potential of multiple LLMs via multi-objective opti...

arXiv 2024
[2025]

Artificial Intelligence (2025), 104308

Stochastic population update can provably be helpful in multi-objective evolutionary algorithms. Artificial Intelligence (2025), 104308. [Bonnabel(2013)] Silv `ere Bonnabel. 2013. Stochastic gradient descent on Riemannian manifolds. IEEE Trans. Automat. Control 58, 9 (2013), 2217–2229. [Brockman et al.(2016)] Greg Brockman, Vicki Cheung, Ludwig Pettersson...

work page 2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2009]

Progress in Aerospace Sciences 45 (2009), 50–79

Recent advances in surrogate-based optimization. Progress in Aerospace Sciences 45 (2009), 50–79. [Francon et al.(2020)] Olivier Francon, Sergio Gonzalez, Babak Hodjat, El- liot Meyerson, Risto Miikkulainen, Xiaoyong Qiu, and Hojjat Shahrzad

work page 2009

[2] [2016]

arXiv preprint arXiv:1606.01540

OpenAI Gym. arXiv preprint arXiv:1606.01540. [Cetin et al.(2022)] Edoardo Cetin, Benjamin Chamberlain, Michael Bron- stein, and Jonathan J. Hunt. 2022. Hyperbolic deep reinforcement learning. arXiv preprint arXiv:2210.01542. [Chrabaszcz et al.(2018)] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. 2018. Back to basics: Benchmarking canonical evoluti...

arXiv 2022

[3] [2020]

In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO)

Effective reinforcement learning through evolutionary surrogate- assisted prescription. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO) . ACM, 814–822. [Fujimoto et al.(2018)] Scott Fujimoto, Herke Hoof, and David Meger. 2018. Addressing function approximation error in actor-critic methods. In International Conference on Mach...

arXiv 2018

[4] [2024]

Advances in Neural Information Processing Systems 36 (2024)

Plastic: Improving input and label plasticity for sample efficient reinforcement learning. Advances in Neural Information Processing Systems 36 (2024). [Li et al.(2024)] Bingdong Li, Zixiang Di, Yanting Yang, Hong Qian, Peng Yang, Hao Hao, Ke Tang, and Aimin Zhou. 2024. It’s morphing time: Unleashing the potential of multiple LLMs via multi-objective opti...

arXiv 2024

[5] [2025]

Artificial Intelligence (2025), 104308

Stochastic population update can provably be helpful in multi-objective evolutionary algorithms. Artificial Intelligence (2025), 104308. [Bonnabel(2013)] Silv `ere Bonnabel. 2013. Stochastic gradient descent on Riemannian manifolds. IEEE Trans. Automat. Control 58, 9 (2013), 2217–2229. [Brockman et al.(2016)] Greg Brockman, Vicki Cheung, Ludwig Pettersson...

work page 2025