Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

How Should We Meta-Learn Reinforcement Learning Algorithms?

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper compares four meta-learning strategies—black-box evolution, black-box distillation, symbolic distillation, and LLM proposal—for discovering reinforcement learning algorithms and distils practical guidelines for choosing among…

desk verdict A genuinely useful first systematic comparison of meta-learning methods for learned RL algorithms, but the method rankings are single-seed stories; send it out and ask for meta-seed replication. read the letter →

arxiv 2507.17668 v2 pith:DZSMDMJM submitted 2025-07-23 cs.LG cs.AI

Alexander David Goldie , Zilin Wang , Jaron Cohen , Jakob Nicolaus Foerster , Shimon Whiteson This is my paper
classification cs.LGcs.AI
keywords meta-learningreinforcementlearninglearnedalgorithmsLLMalgorithmdiscoveryknowledgedistillationsymbolicregressionevolutionstrategiesoptimizers
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

This paper asks which meta-learning algorithm researchers should use when discovering reinforcement learning algorithms, and answers with a head-to-head comparison of black-box evolution, black-box distillation, symbolic distillation, and LLM proposal across four learned components: a PPO drift function (LPO), a recurrent policy-update rule (LPG), and learned optimisers with few and many inputs (No Features and OPEN). Its central findings are three empirical guidelines: language models find effective RL algorithms sample-efficiently when search is warm-started from a performant handcrafted algorithm, though usually with weaker in-distribution performance than black-box learning; distilling a learned algorithm into a same-sized network sometimes improves generalisation at no extra sample cost; and symbolic representations do not scale to recurrent algorithms or algorithms with many inputs. The paper argues these guidelines should steer future meta-learning toward LLM proposal for simple, interpretable components and toward black-box learning for high-dimensional, long-unroll settings.

What carries the argument

The comparison is carried by a suite of target algorithms chosen to vary along three axes: number of inputs, recurrent versus feed-forward structure, and which part of the RL pipeline they replace. They are LPO, a seven-input drift function for PPO; LPG, a recurrent policy-update rule with nineteen inputs trained by meta-gradients; OPEN, an optimiser conditioning on twenty features including neuron dormancy and training progress; and No Features, an eight-input optimiser. Against these, the paper runs four meta-learning mechanisms: evolution strategies over black-box networks; black-box distillation, in which a student network is fit to the teacher's outputs by L2 regression on synthetically generated inputs with no extra environment samples; symbolic distillation, in which the PySR package evolves abstract-syntax-tree programs to match the teacher; and LLM proposal, in which the language model GPT o3-mini iterates on code proposals with in-context fitness feedback, warm-started from a handcrafted algorithm. The load-bearing mechanism for the LLM results is the warm-start initialisation together with per-environment learning-rate tuning: removing either degrades or disables that method, as the paper notes when it cannot apply LLM proposal to LPG. Results are aggregated as interquartile-mean returns with stratified bootstrap confidence intervals over sixteen environment seeds.

What would settle it

A replication that meta-trains each method under comparison for several independent seeds, with identical budgets, and checks whether LLM proposal's advantage and distillation's gains persist across seeds, would settle whether the reported guidelines describe the methods or their single runs.

Watch

Extended reading notes

Core claim

The paper's central claim is that the choice of meta-learning algorithm, not just the choice of learned algorithm, determines whether meta-learning RL algorithms is practical, and that this choice can be guided by a few properties of the target component. Concretely, it reports that LLM-driven proposal, warm-started from a handcrafted algorithm such as PPO or Adam, is the most sample-efficient method studied and produces the most generalisable and interpretable algorithms, provided the component has few inputs the LLM can understand and a per-environment learning rate can be tuned at meta-test time. It reports that distilling a black-box learned algorithm into a same-size network often improves out-of-distribution generalisation without extra environment samples, while smaller distillation frequently hurts. It reports that symbolic distillation, despite yielding interpretable functions, fails to scale: it degrades on the eight-input No Features optimiser, collapses to near-constant functions on OPEN's twenty inputs, and is infeasible for recurrent algorithms with long unrolls such as LPG and recurrent OPEN. On this basis the paper recommends prompting an LLM when a warm-start exists and inputs are interpretable, preferring LLM proposal over symbolic distillation, trying same-size black-box distillation as a cheap regulariser, and relying on black-box learning when input dimensionality or unroll length is large.

Load-bearing premise

The rankings assume that one meta-training run per method, without meta-hyperparameter tuning, as the paper states in Section 6, is representative of that method, so a lucky or unlucky single run could drive the observed differences.

Editorial extensions

If this is right

  • Researchers meta-learning simple, interpretable RL components with few inputs should warm-start an LLM proposer and budget for meta-test-time learning-rate tuning, since this yields strong generalisation with far fewer environment samples than black-box learning.
  • For any feed-forward or short-rollout algorithm learned as a black box, training a same-size distilled copy is worth trying: it adds no environment samples and often improves out-of-distribution generalisation, while a smaller student can reduce performance.
  • Symbolic distillation should not be the default route to interpretable learned RL algorithms: at eight or more inputs it degrades, and for recurrent or long-unroll algorithms it is either infeasible or poor, so LLM proposal is the better interpretability route when warm-starting is possible.
  • When the learned component has many inputs or long recurrent unrolls, as with OPEN, black-box learning remains the only scalable meta-learner, so practitioners should expect high sample costs there rather than trying symbolic or LLM shortcuts.
  • LLM proposal requires a performant initialisation function to exist; for components with no such baseline, which the paper notes for LPG, LLM proposal is not directly applicable.
  • LLM-proposed algorithms usually underperform black-box learning in-distribution, so users who prioritise peak training performance over sample cost and interpretability should still prefer the black-box route.
  • A natural test the paper does not run: meta-training each method over several seeds would reveal whether the reported rankings, such as LLM proposal beating black-box learning on Ant and MinAtar, are stable or an artifact of the paper's single-seed protocol.
  • Blending meta-learners, which the paper only floats as future work, is the natural next lever: warm-starting LLM proposal from a symbolically distilled function, or feeding black-box-encoded inputs into symbolic search, could combine the sample efficiency and interpretability observed separately.

Reading between the lines

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

  • Not run in the paper: meta-training each method over several seeds would reveal whether the reported rankings, such as LLM proposal beating black-box learning on Ant and MinAtar, are stable or an artifact of the paper's single-seed protocol.
  • Blending meta-learners, which the paper only floats as future work, is the natural next lever: warm-starting LLM proposal from a symbolically distilled function, or feeding black-box-encoded inputs into symbolic search, could combine the sample efficiency and interpretability observed separately.
  • LLM proposal's edge likely reflects how well gradient-based optimisation is represented in the model's training corpus; components that stray from familiar optimiser and loss families would probably narrow the gap to black-box learning, sharpening the boundary of the paper's first guideline.
  • The regularising effect of same-size distillation parallels self-distillation results in supervised learning, suggesting the gains might compound when distillation is applied repeatedly or when the teacher is an ensemble, which would be a testable extension.
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.

Referee Report

3 major / 4 minor

Summary. The paper presents an empirical comparison of four meta-learning algorithms for discovering RL algorithms: black-box evolution, same-size and smaller black-box distillation, symbolic distillation, and LLM-based proposal. These are applied to several learned components (LPO, a No-Features learned optimiser, feed-forward and recurrent OPEN, and LPG) and evaluated on in-distribution and out-of-distribution returns, sample cost, runtime, interpretability, and scalability. The central claims are that LLM proposal is sample-efficient when warm-started from a strong handcrafted algorithm, that same-size black-box distillation sometimes improves generalisation without extra environment samples, and that symbolic distillation does not scale to recurrent or many-input algorithms. The paper concludes with a set of design recommendations for meta-learning RL algorithms.

Significance. If the findings hold, this is a useful and timely contribution: the community lacks direct comparisons of meta-learning algorithms for RL, and the proposed guidelines could save substantial compute and guide method choice. The paper is unusually transparent for an empirical study: all hyperparameters are in Supplementary A, discovered functions and LLM prompts are in Supplementary C/D, an example LLM conversation is included, and the limitations of single-seed meta-training are explicitly acknowledged. The use of IQM with bootstrap confidence intervals and a diverse set of external environments is also a strength. However, the central claims are empirical guidelines whose statistical basis is weakened by the single-seed meta-training protocol; the paper's significance is therefore conditional on that limitation being addressed or the claims being appropriately reframed.

major comments (3)
  1. [Section 6, "Due to the high cost of meta-learning ..." and Figures 2-6] The paper meta-trains each algorithm for a single seed and without meta-hyperparameter tuning, yet the reported IQM confidence intervals are over 16 environment seeds. These intervals therefore measure environment-level stochasticity conditional on one meta-training draw, not the variability of the meta-learning method itself. This is load-bearing because several central recommendations, e.g., LLM proposal being "by far the strongest baseline" for No Features in Ant (Section 7.2) and symbolic distillation collapsing on OPEN (Section 7.3), are comparisons between individual stochastic runs of ES, one LLM conversation, and one PySR search. A different meta-training seed could plausibly change the rankings. The paper should either run multiple meta-training seeds for at least the main comparisons, or provide evidence that meta-training seed variance is small relative to the reported effects; otherwise the guidelines should be explicitly reframed as case-study observations rather than robust method comparisons.
  2. [Section 6 (hyperparameter specification) and Section 7.2] LLM proposals are given per-environment learning-rate tuning for their warm-start initialisation, and checkpoint selection is done by in-distribution return, whereas black-box learning is explicitly not meta-hyperparameter-tuned. This asymmetry creates an unequal comparison: the sample-efficiency and performance advantages attributed to LLM proposal may partly reflect the extra tuning budget and meta-test-time evaluation, not the LLM's search ability. The paper acknowledges the learning-rate dependence qualitatively, but the guidelines in Section 8, especially the first bullet, would be more defensible if the comparison included a tuned black-box baseline or a sensitivity analysis showing that the qualitative conclusions are robust to the tuning budget.
  3. [Section 7.2/7.3, Figures 3 and 4] Symbolic distillation results are omitted from the plots when the discovered optimiser produces NaN returns out of distribution, with only a brief mention in the captions. Since one of the central claims is that symbolic representations do not scale well to recurrent or many-input algorithms, the paper should quantify how often symbolic distillation produced NaN or otherwise invalid behaviour, and report the failure rate alongside the IQM of the non-NaN runs. Omitting failed runs entirely makes the aggregate plots difficult to interpret and weakens the very scalability claim the paper wants to support.
minor comments (4)
  1. [Introduction, first paragraph] The text states "We provide our code in an open-source library for meta-learning algorithms" but no repository name, URL, or citation is given; please add a link or a footnote.
  2. [Figure 2 caption] Typo: "algroithm" should be "algorithm".
  3. [Supplementary F captions] Typo: "envrionments" should be "environments"; also "gridwords" should be "gridworlds" in the caption of Figure 15.
  4. [Section 7.4] The phrase "it needssomethingto start from" appears to be a formatting error; insert spaces for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper's guidelines are empirical comparisons against external RL benchmarks; self-citations appear only as baselines and are not load-bearing proof.

full rationale

The paper makes empirical claims about meta-learning algorithms and does not present a formal derivation in which a prediction reduces to its inputs. LLM proposal is explicitly conditional on a warm-start ('so long as there is a good algorithm from which to kickstart meta-training'), and the paper acknowledges that its good out-of-distribution results are partly inherited from the warm-start PPO or Adam and from per-environment tuned learning rates. Those are disclosed design choices, not hidden fitted parameters renamed as predictions. Distillation results are measured by evaluating students on external environments after checkpoint selection on in-distribution return, so the reported generalisation gains are not forced by the selection procedure. The exclusion of symbolic distillation for recurrent algorithms is argued from input dimensionality before any conclusion is drawn, and the empirical failures of symbolic distillation on OPEN are measured rather than assumed. The only self-citations (e.g., Goldie et al. 2024 for OPEN, No Features, and the pretrained recurrent OPEN optimizer) provide external baselines and prior architecture choices; they do not define the paper's conclusions. Single-seed meta-training is a statistical weakness that limits the strength of the rankings, but it is not circularity: no equation equates an output to an input by construction, and no fitted value is renamed as a prediction. Therefore the appropriate circularity score is 0.

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

No new physical or mathematical entities are introduced; the study is empirical. The central claims rest on hyperparameter choices (learning rates, ES settings, distillation budgets, LLM prompt and model), on the mirror-learning drift assumptions, and on the representativeness of the selected algorithms and single-seed meta-training. These are all disclosed in the text, but they limit how far the recommendations can be extrapolated.

free parameters (5)
  • per-environment optimizer learning rates = e.g., SGD LR 0.52 for MinAtar Breakout, 0.17 for Ant; L2O Adam LR 3e-3 for Breakout, 3e-4 for Ant (Tables 4-6)
    LLM-proposed optimizers are evaluated with a learning rate tuned for their warm-start algorithm in each environment, which affects the comparison.
  • ES meta-learning hyperparameters = ES LR 3e-2, sigma init 3e-2, sigma decay 0.999, population size 64 (Table 8)
    Black-box learning performance depends on these evolution strategy settings, which were not meta-hyperparameter tuned.
  • distillation learning rate = sweep over [0.1, 0.02, 0.001], best checkpoint selected by in-distribution return (Appendix A.3)
    Distillation results depend on this sweep and on checkpoint selection by i.d. performance.
  • symbolic distillation budgets = Max Size 40 or 60, Populations 31 or 160, 400 effective PySR iterations (Table 10)
    Symbolic search performance depends on these budgets and on choosing the lowest L2 loss function rather than a complexity-regularized one.
  • LLM model and prompt = GPT o3-mini with DiscoPOP-style prompts (Sections 4.5 and Appendix D)
    LLM proposal results are tied to the specific proprietary model and prompt design; no version pinning is provided.
assumptions (5)
  • standard math Markov decision process formalism and discounted-return objective
    Section 3 defines RL as an MDP; all experiments assume this formalism.
  • domain assumption Mirror learning drift validity conditions (nonnegative, zero at identity, zero gradient at identity)
    Section 3 states valid drift functions must satisfy these; LPO and LLM proposals are constrained by them.
  • ad hoc to paper Single-seed meta-training is representative of each meta-learning algorithm
    Section 6 meta-learns each algorithm for one seed without meta-hyperparameter tuning; the paper's rankings assume this seed is sufficient.
  • domain assumption The selected learned algorithms cover the relevant design space
    Section 5 chooses LPO, No Features, OPEN, and LPG to span inputs, recurrence, and pipeline component; the guidelines assume this coverage generalizes.
  • domain assumption LLM pretraining contains enough optimization knowledge to propose performant algorithms
    LLM proposal assumes GPT o3-mini can generate reasonable code from prompts; no independent verification of this knowledge is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Should We Meta-Learn Reinforcement Learning Algorithms?." pith.science (2026). https://pith.science/paper/DZSMDMJM

@misc{pith2026250717668,
  author       = {Pith},
  title        = {Pith review of: How Should We Meta-Learn Reinforcement Learning Algorithms?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZSMDMJM}},
  note         = {Machine review of arXiv:2507.17668}
}
read the original abstract

The process of meta-learning algorithms from data, instead of relying on manual design, is growing in popularity as a paradigm for improving the performance of machine learning systems. Meta-learning shows particular promise for reinforcement learning (RL), where algorithms are often adapted from supervised or unsupervised learning despite their suboptimality for RL. However, until now there has been a severe lack of comparison between different meta-learning algorithms, such as using evolution to optimise over black-box functions or LLMs to propose code. In this paper, we carry out this empirical comparison of the different approaches when applied to a range of meta-learned algorithms which target different parts of the RL pipeline. In addition to meta-train and meta-test performance, we also investigate factors including the interpretability, sample cost and train time for each meta-learning algorithm. Based on these findings, we propose several guidelines for meta-learning new RL algorithms which will help ensure that future learned algorithms are as performant as possible.

Figures

Figures reproduced from arXiv: 2507.17668 by the authors.

Figure 1
Figure 1. In the RL training loop on the left, we show different components of reinforcement learn [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. IQM of final returns for LPO trained on Ant (top) and MinAtar (bottom). The distillation [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. IQM of final returns for the No Features optimiser after meta-training in Ant (top) and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: IQM of final returns for meta-training feed-forward OPEN in Ant (top) and MinAtar [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: O.o.d. performance of LPG after training in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: IQM of final returns for recurrent OPEN meta-trained in MinAtar. Distillation struggles [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: IQM of final meta-test returns after meta-training LPO in Ant. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: IQM of final meta-test returns for LPO meta-trained in MinAtar. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: IQM of final returns after meta-training the No Features optimizer in Ant. [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: IQM of final returns from meta-training the No Features optimizer in MinAtar. [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: IQM of final returns for meta-training feed-forward OPEN in Ant. [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: IQM of final returns for meta-training feed-forward OPEN in MinAtar. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]
Figure 14
Figure 14. Figure 14: IQM of final returns for meta-training recurrent OPEN in MinAtar. [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: IQM of final returns for LPO after training on gridwords. Results are aggregated across [PITH_FULL_IMAGE:figures/full_fig_p042_15.png]
Figure 16
Figure 16. Figure 16: IQM of final returns on o.o.d. envrionments for the No Feature optimiser after meta [PITH_FULL_IMAGE:figures/full_fig_p042_16.png]
Figure 17
Figure 17. Figure 17: IQM of final returns on o.o.d. envrionments for Feed-Forward OPEN after meta-training [PITH_FULL_IMAGE:figures/full_fig_p042_17.png]
Figure 18
Figure 18. Figure 18: Visualisation of gradients for LPO meta-trained in Ant. [PITH_FULL_IMAGE:figures/full_fig_p043_18.png]
Figure 19
Figure 19. Figure 19: Visualisation of gradients for LPO meta-trained in MinAtar. [PITH_FULL_IMAGE:figures/full_fig_p043_19.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. LEMUR: Learning to Align with Multi-Objective Reinforcement Learning from Preference Feedback

    cs.AI 2026-07 conditional novelty 5.0 of 10

    LEMUR jointly learns a separate reward model for each teacher's preferences and uses them to train a population of multi-objective policies, beating baselines that merge feedback into one reward.

Reference graph

Works this paper leans on

85 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    Zaheer Abbas, Rosie Zhao, Joseph Modayil, Adam White, and Marlos C. Machado. Loss of Plasticity in Continual Deep Reinforcement Learning , March 2023. URL http://arxiv.org/abs/2303.07507

    arXiv 2023

  2. [2]

    Towards characterizing divergence in deep q-learning

    Joshua Achiam, Ethan Knight, and Pieter Abbeel. Towards characterizing divergence in deep q-learning. arXiv preprint arXiv:1903.08894, 2019

    arXiv 1903

  3. [3]

    A Method for Evaluating Hyperparameter Sensitivity in Reinforcement Learning , February 2025

    Jacob Adkins, Michael Bowling, and Adam White. A Method for Evaluating Hyperparameter Sensitivity in Reinforcement Learning , February 2025. URL http://arxiv.org/abs/2412.07165. arXiv:2412.07165 [cs]

    arXiv 2025

  4. [4]

    Deep reinforcement learning at the edge of the statistical precipice

    Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc G Bellemare. Deep reinforcement learning at the edge of the statistical precipice. Advances in Neural Information Processing Systems, 2021

    2021

  5. [5]

    A Generalizable Approach to Learning Optimizers , June 2021

    Diogo Almeida, Clemens Winter, Jie Tang, and Wojciech Zaremba. A Generalizable Approach to Learning Optimizers , June 2021. URL http://arxiv.org/abs/2106.00958

    arXiv 2021

  6. [6]

    Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando de Freitas

    Marcin Andrychowicz, Misha Denil, Sergio Gómez Colmenarejo, Matthew W. Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando de Freitas. Learning to learn by gradient descent by gradient descent. In Proceedings of the 30th international conference on neural information processing systems, NIPS '16, pp.\ 3988--3996, Red Hook, NY, USA, 2016. Curr...

    2016

  7. [7]

    An information-theoretic perspective on intrinsic motivation in reinforcement learning: A survey

    Arthur Aubret, Laetitia Matignon, and Salima Hassas. An information-theoretic perspective on intrinsic motivation in reinforcement learning: A survey. Entropy, 25 0 (2): 0 327, 2023

    2023

  8. [8]

    A Survey of Meta - Reinforcement Learning , August 2024

    Jacob Beck, Risto Vuorio, Evan Zheran Liu, Zheng Xiong, Luisa Zintgraf, Chelsea Finn, and Shimon Whiteson. A Survey of Meta - Reinforcement Learning , August 2024. URL http://arxiv.org/abs/2301.08028

    arXiv 2024

Show all 85 references
  1. [9]

    OpenAI gym, 2016

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. OpenAI gym, 2016

    2016

  2. [10]

    Exploration by random network distillation

    Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation. arXiv preprint arXiv:1810.12894, 2018

    2018 arXiv

  3. [11]

    Boltzmann exploration done right

    Nicol \`o Cesa-Bianchi, Claudio Gentile, G \'a bor Lugosi, and Gergely Neu. Boltzmann exploration done right. Advances in neural information processing systems, 30, 2017

    2017

  4. [12]

    Xiangning Chen, Chen Liang, Da Huang, Esteban Real, Kaiyuan Wang, Yao Liu, Hieu Pham, Xuanyi Dong, Thang Luong, Cho-Jui Hsieh, Yifeng Lu, and Quoc V. Le. Symbolic Discovery of Optimization Algorithms , May 2023. URL http://arxiv.org/abs/2302.06675

    2023 arXiv

  5. [13]

    Interpretable Machine Learning for Science with PySR and SymbolicRegression .jl, May 2023

    Miles Cranmer. Interpretable Machine Learning for Science with PySR and SymbolicRegression .jl, May 2023. URL http://arxiv.org/abs/2305.01582. arXiv:2305.01582 [astro-ph]

    2023 arXiv

  6. [14]

    Discovering Symbolic Models from Deep Learning with Inductive Biases , November 2020

    Miles Cranmer, Alvaro Sanchez-Gonzalez, Peter Battaglia, Rui Xu, Kyle Cranmer, David Spergel, and Shirley Ho. Discovering Symbolic Models from Deep Learning with Inductive Biases , November 2020. URL http://arxiv.org/abs/2006.11287

    2020 arXiv

  7. [15]

    DeepSeek-AI , Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei ...

    2025 arXiv

  8. [16]

    Emergent Complexity and Zero -shot Transfer via Unsupervised Environment Design , February 2021

    Michael Dennis, Natasha Jaques, Eugene Vinitsky, Alexandre Bayen, Stuart Russell, Andrew Critch, and Sergey Levine. Emergent Complexity and Zero -shot Transfer via Unsupervised Environment Design , February 2021. URL http://arxiv.org/abs/2012.02096

    2021 arXiv

  9. [17]

    Loss of plasticity in deep continual learning

    Shibhansh Dohare, J Fernando Hernandez-Garcia, Qingfeng Lan, Parash Rahman, A Rupam Mahmood, and Richard S Sutton. Loss of plasticity in deep continual learning. Nature, 632 0 (8026): 0 768--774, 2024

    2024

  10. [18]

    Rl ^2 : Fast reinforcement learning via slow reinforcement learning

    Yan Duan, John Schulman, Xi Chen, Peter L Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl ^2 : Fast reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779, 2016

    2016 arXiv

  11. [19]

    Jackson, Andrei Lupu, Alexander D

    Benjamin Ellis, Matthew T. Jackson, Andrei Lupu, Alexander D. Goldie, Mattie Fellows, Shimon Whiteson, and Jakob Foerster. Adam on Local Time : Addressing Nonstationarity in RL with Relative Adam Timesteps , December 2024. URL http://arxiv.org/abs/2412.17113. arXiv:2412.17113 [cs]

    2024 arXiv

  12. [20]

    OMNI - EPIC : Open -endedness via Models of human Notions of Interestingness with Environments Programmed in Code , May 2024

    Maxence Faldor, Jenny Zhang, Antoine Cully, and Jeff Clune. OMNI - EPIC : Open -endedness via Models of human Notions of Interestingness with Environments Programmed in Code , May 2024. URL http://arxiv.org/abs/2405.15568. arXiv:2405.15568 [cs]

    2024 arXiv

  13. [21]

    Model-agnostic meta-learning for fast adaptation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pp.\ 1126--1135. PMLR, 2017

    2017

  14. [22]

    Noisy Networks for Exploration , July 2019

    Meire Fortunato, Mohammad Gheshlaghi Azar, Bilal Piot, Jacob Menick, Ian Osband, Alex Graves, Vlad Mnih, Remi Munos, Demis Hassabis, Olivier Pietquin, Charles Blundell, and Shane Legg. Noisy Networks for Exploration , July 2019. URL http://arxiv.org/abs/1706.10295. arXiv:1706....

    2019 arXiv

  15. [23]

    Daniel Freeman, Erik Frey, Anton Raichuk, Sertan Girgin, Igor Mordatch, and Olivier Bachem

    C. Daniel Freeman, Erik Frey, Anton Raichuk, Sertan Girgin, Igor Mordatch, and Olivier Bachem. Brax - a differentiable physics engine for large scale rigid body simulation, 2021. URL http://github.com/google/brax

    2021

  16. [24]

    Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar

    Tommaso Furlanello, Zachary C. Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born Again Neural Networks , June 2018. URL http://arxiv.org/abs/1805.04770. arXiv:1805.04770 [stat]

    2018 arXiv

  17. [25]

    Goldie, Chris Lu, Matthew T

    Alexander D. Goldie, Chris Lu, Matthew T. Jackson, Shimon Whiteson, and Jakob N. Foerster. Can Learned Optimization Make Reinforcement Learning Less Difficult ? In Advances in Neural Information Processing Systems , volume 37, pp.\ 5454--5497, 2024

    2024

  18. [26]

    Benchmarking the spectrum of agent capabilities

    Danijar Hafner. Benchmarking the spectrum of agent capabilities. arXiv preprint arXiv:2109.06780, 2021

    2021 arXiv

  19. [27]

    Distilling the Knowledge in a Neural Network , March 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network , March 2015. URL http://arxiv.org/abs/1503.02531. arXiv:1503.02531 [stat]

    2015 arXiv

  20. [28]

    Long short-term memory

    Sepp Hochreiter and J\" u rgen Schmidhuber. Long short-term memory. Neural Comput., 9 0 (8): 0 1735–1780, November 1997. ISSN 0899-7667. doi:10.1162/neco.1997.9.8.1735. URL https://doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

  21. [29]

    Automated Design of Agentic Systems , August 2024

    Shengran Hu, Cong Lu, and Jeff Clune. Automated Design of Agentic Systems , August 2024. URL http://arxiv.org/abs/2408.08435

    2024 arXiv

  22. [30]

    Transient non-stationarity and generalisation in deep reinforcement learning

    Maximilian Igl, Gregory Farquhar, Jelena Luketina, Wendelin Boehmer, and Shimon Whiteson. Transient non-stationarity and generalisation in deep reinforcement learning. In International Conference on Learning Representations, 2021 a . URL https://openreview.net/forum?id=Qun8fv4qSby

    2021

  23. [31]

    Transient non-stationarity and generalisation in deep reinforcement learning

    Maximilian Igl, Gregory Farquhar, Jelena Luketina, Wendelin Boehmer, and Shimon Whiteson. Transient non-stationarity and generalisation in deep reinforcement learning. In International conference on learning representations, 2021 b . URL https://openreview.net/forum?id=Qun8fv4qSby

    2021

  24. [32]

    Discovering General Reinforcement Learning Algorithms with Adversarial Environment Design , October 2023

    Matthew Thomas Jackson, Minqi Jiang, Jack Parker-Holder, Risto Vuorio, Chris Lu, Gregory Farquhar, Shimon Whiteson, and Jakob Nicolaus Foerster. Discovering General Reinforcement Learning Algorithms with Adversarial Environment Design , October 2023. URL http://arxiv.org/abs/2...

    2023 arXiv

  25. [33]

    Discovering temporally-aware reinforcement learning algorithms

    Matthew Thomas Jackson, Chris Lu, Louis Kirsch, Robert Tjarko Lange, Shimon Whiteson, and Jakob Nicolaus Foerster. Discovering temporally-aware reinforcement learning algorithms. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview....

    2024

  26. [34]

    Improving policy optimization with generalist-specialist learning

    Zhiwei Jia, Xuanlin Li, Zhan Ling, Shuang Liu, Yiran Wu, and Hao Su. Improving policy optimization with generalist-specialist learning. In International Conference on Machine Learning, pp.\ 10104--10119. PMLR, 2022

    2022

  27. [35]

    Meta Learning Backpropagation And Improving It , March 2022

    Louis Kirsch and Jürgen Schmidhuber. Meta Learning Backpropagation And Improving It , March 2022. URL http://arxiv.org/abs/2012.14905

    2022 arXiv

  28. [36]

    Improving generalization in meta reinforcement learning using learned objectives

    Louis Kirsch, Sjoerd van Steenkiste, and Juergen Schmidhuber. Improving generalization in meta reinforcement learning using learned objectives. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=S1evHerYPr

    2020

  29. [37]

    Mirror Learning : A Unifying Framework of Policy Optimisation , November 2024

    Jakub Grudzien Kuba, Christian Schroeder de Witt, and Jakob Foerster. Mirror Learning : A Unifying Framework of Policy Optimisation , November 2024. URL http://arxiv.org/abs/2201.02373. arXiv:2201.02373 [cs]

    2024 arXiv

  30. [38]

    Rupam Mahmood, Shuicheng Yan, and Zhongwen Xu

    Qingfeng Lan, A. Rupam Mahmood, Shuicheng Yan, and Zhongwen Xu. Learning to Optimize for Reinforcement Learning , June 2024. URL http://arxiv.org/abs/2302.01470

    2024 arXiv

  31. [39]

    gymnax: A JAX -based reinforcement learning environment library, 2022 a

    Robert Tjarko Lange. gymnax: A JAX -based reinforcement learning environment library, 2022 a . URL http://github.com/RobertTLange/gymnax

    2022

  32. [40]

    evosax: JAX -based Evolution Strategies

    Robert Tjarko Lange. evosax: JAX -based Evolution Strategies . arXiv preprint arXiv:2212.04180, 2022 b

    2022 arXiv

  33. [41]

    In-context reinforcement learning with algorithm distillation

    Michael Laskin, Luyu Wang, Junhyuk Oh, Emilio Parisotto, Stephen Spencer, Richie Steigerwald, DJ Strouse, Steven Stenberg Hansen, Angelos Filos, Ethan Brooks, maxime gazeau, Himanshu Sahni, Satinder Singh, and Volodymyr Mnih. In-context reinforcement learning with algorithm di...

    2023

  34. [42]

    Joel Lehman, Jonathan Gordon, Shawn Jain, Kamal Ndousse, Cathy Yeh, and Kenneth O. Stanley. Evolution through Large Models , June 2022. URL http://arxiv.org/abs/2206.08896

    2022 arXiv

  35. [43]

    Rediscovering orbital mechanics with machine learning

    Pablo Lemos, Niall Jeffrey, Miles Cranmer, Shirley Ho, and Peter Battaglia. Rediscovering orbital mechanics with machine learning. Machine Learning: Science and Technology, 4 0 (4): 0 045002, 2023

    2023

  36. [44]

    Discovered policy optimisation

    Chris Lu, Jakub Kuba, Alistair Letcher, Luke Metz, Christian Schroeder de Witt, and Jakob Foerster. Discovered policy optimisation. Advances in Neural Information Processing Systems, 35: 0 16455--16468, 2022

    2022

  37. [45]

    Chan, Jakob Foerster, Mihaela van der Schaar, and Robert Tjarko Lange

    Chris Lu, Samuel Holt, Claudio Fanconi, Alex J. Chan, Jakob Foerster, Mihaela van der Schaar, and Robert Tjarko Lange. Discovering Preference Optimization Algorithms with and for Large Language Models , September 2024. URL http://arxiv.org/abs/2406.08414

    2024 arXiv

  38. [46]

    Behaviour Distillation

    Andrei Lupu, Chris Lu, Jarek Luca Liesen, Robert Tjarko Lange, and Jakob Nicolaus Foerster. Behaviour Distillation . In The Twelfth International Conference on Learning Representations , 2024. URL https://openreview.net/forum?id=qup9xD8mW4

    2024

  39. [47]

    Understanding plasticity in neural networks, August 2023

    Clare Lyle, Zeyu Zheng, Evgenii Nikishin, Bernardo Avila Pires, Razvan Pascanu, and Will Dabney. Understanding plasticity in neural networks, August 2023. URL http://arxiv.org/abs/2303.01486

    2023 arXiv

  40. [48]

    Craftax: a lightning-fast benchmark for open-ended reinforcement learning

    Michael Matthews, Michael Beukman, Benjamin Ellis, Mikayel Samvelyan, Matthew Jackson, Samuel Coward, and Jakob Foerster. Craftax: a lightning-fast benchmark for open-ended reinforcement learning. In International conference on machine learning ( ICML ) , 2024

    2024

  41. [49]

    Interpretable machine learning methods applied to jet background subtraction in heavy-ion collisions

    Tanner Mengel, Patrick Steffanic, Charles Hughes, Antonio Carlos Oliveira da Silva, and Christine Nattrass. Interpretable machine learning methods applied to jet background subtraction in heavy-ion collisions. Physical Review C, 108 0 (2): 0 L021901, 2023

    2023

  42. [50]

    Meta- Learning Update Rules for Unsupervised Representation Learning , February 2019 a

    Luke Metz, Niru Maheswaranathan, Brian Cheung, and Jascha Sohl-Dickstein. Meta- Learning Update Rules for Unsupervised Representation Learning , February 2019 a . URL http://arxiv.org/abs/1804.00222. arXiv:1804.00222 [cs, stat]

    2019 arXiv

  43. [51]

    Daniel Freeman, and Jascha Sohl-Dickstein

    Luke Metz, Niru Maheswaranathan, Jeremy Nixon, C. Daniel Freeman, and Jascha Sohl-Dickstein. Understanding and correcting pathologies in the training of learned optimizers, June 2019 b . URL http://arxiv.org/abs/1810.10180

    2019 arXiv

  44. [52]

    Daniel Freeman, Ben Poole, and Jascha Sohl-Dickstein

    Luke Metz, Niru Maheswaranathan, C. Daniel Freeman, Ben Poole, and Jascha Sohl-Dickstein. Tasks, stability, architecture, and compute: Training more effective learned optimizers, and using them to train themselves, September 2020. URL http://arxiv.org/abs/2009.11243

    2020 arXiv

  45. [53]

    Daniel Freeman, Samuel S

    Luke Metz, C. Daniel Freeman, Samuel S. Schoenholz, and Tal Kachman. Gradients are Not All You Need , January 2022 a . URL http://arxiv.org/abs/2111.05803. arXiv:2111.05803 [cs]

    2022 arXiv

  46. [54]

    Daniel Freeman, Amil Merchant, Lucas Beyer, James Bradbury, Naman Agrawal, Ben Poole, Igor Mordatch, Adam Roberts, and Jascha Sohl-Dickstein

    Luke Metz, James Harrison, C. Daniel Freeman, Amil Merchant, Lucas Beyer, James Bradbury, Naman Agrawal, Ben Poole, Igor Mordatch, Adam Roberts, and Jascha Sohl-Dickstein. VeLO : Training Versatile Learned Optimizers by Scaling Up , November 2022 b . URL http://arxiv.org/abs/2...

    2022 arXiv

  47. [55]

    Self-distillation amplifies regularization in hilbert space

    Hossein Mobahi, Mehrdad Farajtabar, and Peter Bartlett. Self-distillation amplifies regularization in hilbert space. Advances in Neural Information Processing Systems, 33: 0 3351--3361, 2020

    2020

  48. [56]

    Small batch deep reinforcement learning

    Johan Obando Ceron, Marc Bellemare, and Pablo Samuel Castro. Small batch deep reinforcement learning. Advances in Neural Information Processing Systems, 36: 0 26003--26024, 2023

    2023

  49. [57]

    Discovering reinforcement learning algorithms

    Junhyuk Oh, Matteo Hessel, Wojciech M Czarnecki, Zhongwen Xu, Hado P van Hasselt, Satinder Singh, and David Silver. Discovering reinforcement learning algorithms. Advances in Neural Information Processing Systems, 33: 0 1060--1070, 2020

    2020

  50. [58]

    Openai o3-mini, January 2025

    OpenAI. Openai o3-mini, January 2025. URL https://openai.com/index/openai-o3-mini/

    2025

  51. [59]

    Stabilizing transformers for reinforcement learning

    Emilio Parisotto, Francis Song, Jack Rae, Razvan Pascanu, Caglar Gulcehre, Siddhant Jayakumar, Max Jaderberg, Raphael Lopez Kaufman, Aidan Clark, Seb Noury, et al. Stabilizing transformers for reinforcement learning. In International conference on machine learning, pp.\ 7487--...

    2020

  52. [60]

    Evolving Curricula with Regret - Based Environment Design

    Jack Parker-Holder, Minqi Jiang, Michael Dennis, Mikayel Samvelyan, Jakob Foerster, Edward Grefenstette, and Tim Rocktäschel. Evolving Curricula with Regret - Based Environment Design . In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sab...

    2022

  53. [61]

    Chen, Xi Chen, Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz

    Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y. Chen, Xi Chen, Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz. Parameter Space Noise for Exploration , January 2018. URL http://arxiv.org/abs/1706.01905. arXiv:1706.01905 [cs]

    2018 arXiv

  54. [62]

    Tunability: Importance of hyperparameters of machine learning algorithms

    Philipp Probst, Anne-Laure Boulesteix, and Bernd Bischl. Tunability: Importance of hyperparameters of machine learning algorithms. Journal of Machine Learning Research, 20 0 (53): 0 1--32, 2019

    2019

  55. [63]

    Evolutionsstrategie : Optimierung technischer systeme nach prinzipien der biologischen evolution

    Ingo Rechenberg. Evolutionsstrategie : Optimierung technischer systeme nach prinzipien der biologischen evolution. 1973. URL https://api.semanticscholar.org/CorpusID:60975248

    1973

  56. [64]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with larg...

    2024 doi

  57. [65]

    Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell

    Andrei A. Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy Distillation , January 2016. URL http://arxiv.org/abs/1511.06295. arXiv:1511.06295 [cs]

    2016 arXiv

  58. [66]

    Evolution Strategies as a Scalable Alternative to Reinforcement Learning , September 2017

    Tim Salimans, Jonathan Ho, Xi Chen, Szymon Sidor, and Ilya Sutskever. Evolution Strategies as a Scalable Alternative to Reinforcement Learning , September 2017. URL http://arxiv.org/abs/1703.03864. arXiv:1703.03864 [cs, stat]

    2017 arXiv

  59. [67]

    Proximal Policy Optimization Algorithms , August 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal Policy Optimization Algorithms , August 2017. URL http://arxiv.org/abs/1707.06347

    2017 arXiv

  60. [68]

    High- Dimensional Continuous Control Using Generalized Advantage Estimation , October 2018

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High- Dimensional Continuous Control Using Generalized Advantage Estimation , October 2018. URL http://arxiv.org/abs/1506.02438

    2018 arXiv

  61. [69]

    The Dormant Neuron Phenomenon in Deep Reinforcement Learning , June 2023

    Ghada Sokar, Rishabh Agarwal, Pablo Samuel Castro, and Utku Evci. The Dormant Neuron Phenomenon in Deep Reinforcement Learning , June 2023. URL http://arxiv.org/abs/2302.12902

    2023 arXiv

  62. [70]

    Distilling Reinforcement Learning Algorithms for In - Context Model - Based Planning , February 2025

    Jaehyeon Son, Soochan Lee, and Gunhee Kim. Distilling Reinforcement Learning Algorithms for In - Context Model - Based Planning , February 2025. URL http://arxiv.org/abs/2502.19009. arXiv:2502.19009 [cs]

    2025 arXiv

  63. [71]

    Generalizable Symbolic Optimizer Learning

    Xiaotian Song, Peng Zeng, Yanan Sun, and Andy Song. Generalizable Symbolic Optimizer Learning . 2024 a

    2024

  64. [72]

    Position: Leverage Foundational Models for Black - Box Optimization , May 2024 b

    Xingyou Song, Yingtao Tian, Robert Tjarko Lange, Chansoo Lee, Yujin Tang, and Yutian Chen. Position: Leverage Foundational Models for Black - Box Optimization , May 2024 b . URL http://arxiv.org/abs/2405.03547

    2024 arXiv

  65. [73]

    Maxinfo RL : Boosting exploration in reinforcement learning through information gain maximization

    Bhavya Sukhija, Stelian Coros, Andreas Krause, Pieter Abbeel, and Carmelo Sferrazza. Maxinfo RL : Boosting exploration in reinforcement learning through information gain maximization. In The Thirteenth International Conference on Learning Representations, 2025. URL https://ope...

    2025

  66. [74]

    Sutton and Andrew Barto

    Richard S. Sutton and Andrew Barto. Reinforcement learning: an introduction. Adaptive computation and machine learning. The MIT Press, Cambridge, Massachusetts London, England, second edition edition, 2020. ISBN 978-0-262-03924-6

    2020

  67. [75]

    Improving deep reinforcement learning by reducing the chain effect of value and policy churn

    Hongyao Tang and Glen Berseth. Improving deep reinforcement learning by reducing the chain effect of value and policy churn. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=cQoAgPBARc

    2024

  68. [76]

    MuJoCo : A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo : A physics engine for model-based control. In 2012 IEEE / RSJ international conference on intelligent robots and systems , pp.\ 5026--5033. IEEE, 2012. doi:10.1109/IROS.2012.6386109

    2012

  69. [77]

    Deep reinforcement learning and the deadly triad

    Hado Van Hasselt, Yotam Doron, Florian Strub, Matteo Hessel, Nicolas Sonnerat, and Joseph Modayil. Deep reinforcement learning and the deadly triad. arXiv preprint arXiv:1812.02648, 2018

    2018 arXiv

  70. [78]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need , August 2023. URL http://arxiv.org/abs/1706.03762. arXiv:1706.03762 [cs]

    2023 arXiv

  71. [79]

    Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A. Efros. Dataset Distillation , February 2020. URL http://arxiv.org/abs/1811.10959. arXiv:1811.10959 [cs]

    2020 arXiv

  72. [80]

    Natural Evolution Strategies , June 2011

    Daan Wierstra, Tom Schaul, Tobias Glasmachers, Yi Sun, and Jürgen Schmidhuber. Natural Evolution Strategies , June 2011. URL http://arxiv.org/abs/1106.4487

    2011 arXiv

  73. [81]

    Understanding short-horizon bias in stochastic meta-optimization

    Yuhuai Wu, Mengye Ren, Renjie Liao, and Roger Grosse. Understanding short-horizon bias in stochastic meta-optimization. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=H1MczcgR-

    2018

  74. [82]

    MinAtar : An atari-inspired testbed for thorough and reproducible reinforcement learning experiments

    Kenny Young and Tian Tian. MinAtar : An atari-inspired testbed for thorough and reproducible reinforcement learning experiments. arXiv preprint arXiv:1903.03176, 2019

    1903 arXiv

  75. [83]

    Self-distillation as instance-specific label smoothing

    Zhilu Zhang and Mert Sabuncu. Self-distillation as instance-specific label smoothing. Advances in Neural Information Processing Systems, 33: 0 2184--2195, 2020

    2020

  76. [84]

    Symbolic Learning to Optimize : Towards Interpretability and Scalability , May 2022

    Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, and Zhangyang Wang. Symbolic Learning to Optimize : Towards Interpretability and Scalability , May 2022. URL http://arxiv.org/abs/2203.06578

    2022 arXiv

  77. [85]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.