Pith. sign in

REVIEW 3 major objections 6 minor 72 references

A meta-learning algorithm that replaces gradients with greedy selection on stochastic dynamical systems.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:30 UTC pith:AMKNFLVT

load-bearing objection A clear, novel meta-learning proposal whose core selection step is statistically flawed and which lacks any real validation; worth engaging for the ideas, but not for the results. the 3 major comments →

arxiv 2607.23925 v1 pith:AMKNFLVT submitted 2026-07-27 cs.LG

Greedy dynamical meta-learning

Aria Yom This is my paper
classification cs.LG
keywords meta-learningdynamical systemsevolution strategieszeroth-order optimizationself-modifying networkstimescalesgradient-free learninggreedy selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that gradient descent cannot produce learning systems because gradients become unstable beyond a predictability horizon set by Lyapunov exponents, while zeroth-order optimizers are crushed by the curse of dimensionality in high-dimensional parameter spaces. The proposed alternative, dynamical systems meta-learning (DSML), encodes the agent itself as a stochastic dynamical system and selects, over a long time horizon, mutant trajectories that perform best. The central claim is that by selecting high-performing agents over long timescales, one is also selecting for agents that generate beneficial variations, preserve learned structures, and recover from damaging mutations — i.e., that learn how to learn. To make this work, the paper separates a mutation timescale µ from an evaluation timescale ν, argues these must be distinct, and gives an adaptive algorithm for tuning the two remaining timescales.

Core claim

The paper's central proposal is that an agent encoded as a stochastic dynamical system x' = D(x) can be meta-trained by greedy evolutionary selection without any gradient signal. The system state x includes both weights and activations, and the fixed stochastic laws of motion D generate mutant trajectories. The learning cycle spawns m mutants from a parent state, simulates each mutant forward for τ steps while accumulating reward evaluations, and crowns the highest-scoring mutant as the next parent. The author claims that over long horizons this selects not just for high immediate reward but for qualities that make a system learnable — beneficial variations, preservation of learned structure

What carries the argument

The central objects are the stochastic dynamical system D(x) with state x, and the two timescales µ (mutation time, generation length) and ν (evaluation horizon). The inner learning cycle C_µτ combines the three parameters n, k, ν into a single timescale τ by simulating each mutant's trajectory for τ steps and evaluating reward at every timestep t > µ — the '1/t = log t trick' — paying only a log factor over the optimal sampling scheme. The outer tuning algorithm updates log µ and log τ using random-direction stochastic approximation on the gain S = E[f(C_µτ(x)) − f(x)]/µ, with log-parameter updates derived from dS = S(d log τ − d log µ). The paper argues that S is concave in √θ for θ ∈ {1/µ

Load-bearing premise

The fixed stochastic laws of motion D(x) must be well-designed enough that mutant trajectories are discernibly different and can escape attractors; if no such hand-crafted dynamics exist, the whole selection cycle produces no signal.

What would settle it

Run Algorithm 1 on a fixed stochastic dynamical system with a known optimum, using the same compute budget as random search; if the greedy cycle does not outperform equal-compute random search across several random seeds, the central selection claim fails. Alternatively, exhibit a reward function f for which the measurement noise diverges over the timescale τ, violating the second condition in Sec. 2.2 and drowning the selection signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If DSML works, learning can proceed over arbitrary time horizons without gradient truncation, because the outer loop optimizes only two parameters.
  • The log-time sampling scheme means the user never needs to hand-tune the evaluation horizon ν; a single τ covers all timescales at polylogarithmic overhead.
  • By forcing all meta-information into the state x, the algorithm eliminates the separate policy parameters of meta-RL, so whatever the agent learns can directly refine its own mutations.
  • The only hyperparameter requiring careful tuning is the learning rate η of the outer loop; σ² can be fixed at about 0.1 across well-behaved landscapes.

Where Pith is reading between the lines

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

  • The separation of µ and ν is a general principle that likely applies to any evolutionary or black-box optimizer whose objective lags an unobserved latent quality; it could be tested on standard evolution strategies by evaluating mutants at a delayed time rather than immediately.
  • The log-time sampling trick suggests a continuous-time generalization: simulate each trajectory with a Poisson rate in logarithmic time, which would make the algorithm's compute distribution exactly scale-invariant across timescales.
  • A decisive test of the central premise would be to apply Algorithm 1 to a concrete high-dimensional stochastic dynamical network and check whether greedy selection over τ steps outperforms equal-compute random search; the paper defers such experiments to future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes dynamical systems meta-learning (DSML), a framework in which an agent is a stochastic dynamical system whose state subsumes its weights and activations. It argues that gradient-based training is unstable over long horizons and that meta-learning should instead proceed by evolutionary selection over mutant trajectories. The inner loop (Alg. 1) introduces two timescales, the mutation time μ and the evaluation time τ, and uses a log-time sampling scheme intended to collapse the three parameters n, k, ν into a single timescale τ. The outer loop (Alg. 2) tunes μ and τ by random-direction stochastic approximation, with analysis in Sec. 2.4 purporting to show that Q has a unique local maximum and that σ²≈0.1 is a universal choice. The only quantitative example is the OU toy model in Fig. 4 illustrating the gap between μ and ν; no end-to-end run of Alg. 1 or Alg. 2 is presented.

Significance. If the framework worked, DSML would be an unusual gradient-free alternative to meta-learning, and the log-time sampling idea is genuinely intriguing. The paper is clearly written and unusually candid about its limitations, and the taffy-map example is an effective illustration of why zeroth-order estimates can outlast first-order estimates in chaotic systems. However, the contribution is currently conceptual rather than established: the central object D is never instantiated, the selection score in Alg. 1 does not implement the stated objective, the tuning analysis in Sec. 2.4 relies on assumptions the paper itself concedes have counterexamples, and no experiments validate either loop of the algorithm. The strengths are the clarity of the framing and the interesting timescale argument, but they do not yet support a publishable algorithmic claim.

major comments (3)
  1. [Sec. 2.2, Alg. 1] The selection score z_i is not an estimator of E[f(y_i(ν))]. For a single trajectory y(t), Alg. 1 computes z_i = Σ_{j=1}^{τ} Σ_{t=μ}^{τ/j} f(y(t)) = Σ_{t=μ}^{τ} floor(τ/t) f(y(t)), which is approximately τ Σ_{t=μ}^{τ} f(y(t))/t. This weights early times by 1/t, so a mutant with a brief early reward burst can outscore a mutant with superior long-term performance at ν. The paper compares only the compute cost of the log-time scheme ('only a factor of log τ') and never shows that the induced ranking equals or even correlates with the ranking under the stated optimal estimator. The toy model of Sec. 2.1 is exactly a regime in which early f values are not informative about f(ν), so this is a central flaw, not an edge case.
  2. [Sec. 2.4, Eqs. (5)–(8)] Eq. (6) is obtained by imposing dQ=0; it is a stationarity condition, not a general expression for the gradient of S. The statement that the right-hand side of Eqn. (6) is 'easily estimated from samples of S' is therefore circular as a basis for the RDSA updates in Eqs. (7)–(8). In addition, the extra (S_+ + S_-) terms in the update rules are not derived from any displayed objective, and the claim that Q has a unique local maximum and that σ²≈0.1 is universal rests on the concavity-in-sqrt(θ) assumption that the manuscript itself concedes has counterexamples (Sec. 2.4, footnote 5). The tuning loop is thus not actually justified as written.
  3. [Sec. 1.4, Discussion] The central object D(x) is never instantiated. Alg. 1 uses a stochastic mutator g both to generate mutants and to advance trajectories, but no definition connects g to the dynamics D or to the mutation distribution g_μ. The paper states that performance will depend on D being well-designed and defers architecture and dynamics to an upcoming publication. Consequently, there is no reproducible instance of the DSML cycle and no end-to-end simulation of Alg. 1 or Alg. 2; Fig. 4 only illustrates the μ/ν gap for an OU toy, not the algorithm's behavior. The Sec. 1.5 claims (points 6–7) that long-timescale selection selects for beneficial variation, preserves learned structure, and recovers from damaging mutations are therefore untested assertions.
minor comments (6)
  1. [Alg. 1] Alg. 1 initializes y_j^i ← x_i and then updates y with g in the same way that mutants are sampled. The relationship among g, D, and the mutation distribution g_μ needs to be specified precisely; as written, the algorithm is not implementable.
  2. [Sec. 2.2] The parenthetical 'we encourage the reader not to think too much about whether these times are discrete or continuous' masks a real issue: Alg. 1 requires integer loop bounds, while the analysis uses continuous timescales. This should be formalized.
  3. [Alg. 2] The choices T=τ/μ and N=mτT are stated without justification, and no integer-rounding or stopping convention is given when μ and τ are real-valued. This matters because the algorithm updates log μ and log τ continuously.
  4. [Sec. 1.2] The taffy-map example supports the general argument against gradients in chaotic systems but is not connected to the DSML algorithm. It would be useful to state explicitly what the example does and does not establish for the proposal.
  5. [Sec. 2.2] The text says Rechenberg's rule gives m≈10, but Alg. 1 hard-codes m=8. Also, the citation to Rechenberg (1973) is vague; a precise statement of the '1/5 rule' or the relevant result would help the reader evaluate the choice.
  6. [Fig. 4] Numerical integration details for the OU system (time step, number of trajectories, discretization of the Wiener process, seed) are omitted, so the illustration is not reproducible.

Circularity Check

0 steps flagged

No circularity: the DSML cycle, log-time sampling, and timescale tuner are defined constructively; no fitted parameter is relabeled as a prediction and no author self-citation is load-bearing.

full rationale

The paper's derivation chain does not reduce any output to its inputs by construction. The central objects — the DSML cycle (Sec. 1.4), Algorithm 1's log-time sampling (Sec. 2.2), and the RDSA timescale updates (Secs. 2.3–2.4) — are defined explicitly, and the tuning equations (5)–(8) follow algebraically from the definitions S = E[f(C_μτ(x))−f(x)]/μ and Q = μS/τ. No external prediction is claimed, no parameter is fitted to a target result and then renamed, and no prior work by the same author is invoked to carry a load-bearing conclusion. The paper openly states its unresolved premises: "We will have more to say about the system architecture and dynamical equations in an upcoming publication" (Sec. 1.4); "we were not able to devise a principled algorithm along these lines" (Sec. 1.4); and "We encourage the reader not to think too much about whether these times are discrete or continuous" (Sec. 2.2). The most serious technical objection — that Algorithm 1's cumulative z_i is a 1/t-weighted sum rather than an estimator of E[f(y(ν))] and is not shown to rank mutants identically — is a correctness/rigor gap in the paper's own comparison, not circularity: the paper never equates z_i with the target mean by definition. Self-adaptation of µ and τ is standard online tuning (explicitly credited to Schwefel, 1981) and is not a circular derivation. Score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 1 invented entities

The framework rests on several unproven domain assumptions about the existence of well-designed fixed dynamics, the observability of latent intelligence through a scalar reward, and the shape of the performance function S. The free parameters are mostly timescales and learning rates that the algorithm tunes, but their initial values and the universal σ² are chosen by hand.

free parameters (5)
  • mu (mutation timescale)
    Central timescale in Algorithms 1 and 2; initial value left unspecified and tuned online.
  • tau (evaluation horizon)
    Second central timescale; initial value unspecified; algorithm performance depends on it.
  • sigma^2 = 0.1
    Perturbation variance for log µ and log τ in the RDSA update; chosen by hand in Sec. 2.4 and claimed universal for well-behaved concave landscapes.
  • eta (learning rate)
    Outer-loop learning rate; the paper says it is the only hyperparameter requiring careful tuning, but gives no default value.
  • m (mutants per generation) = 8
    Number of mutants per generation, taken from Rechenberg's rule; a chosen constant rather than fitted to a target.
axioms (6)
  • domain assumption The dynamics x'=D(x) are fixed, stochastic, and well-designed; the agent cannot modify them.
    Sec. 1.4: 'Our agents have no agency to rewrite these laws... their performance will likely depend on these laws being well-designed.' If D is poor, DSML fails.
  • domain assumption Latent intelligence is reflected in a scalar reward f at a delayed evaluation timescale ν different from µ.
    Sec. 2.1: 'The hope is that it may suffice to find a good reward function f... there are two distinct timescales.' This is the basis for the entire cycle.
  • domain assumption Measurement error does not vary too much with timescale, and evaluations near ν suffice to distinguish mutants.
    Sec. 2.2 gives this as a condition for Algorithm 1; if noise diverges, the log-all-timescales sampling loses signal.
  • ad hoc to paper Simulating one timestep costs the same as one f evaluation.
    Sec. 2.2: 'let us assume that simulating a single timestep costs the same amount of compute as performing a single f evaluation.' Used to claim only log τ overhead.
  • ad hoc to paper S(µ,τ) is monotone and concave in sqrt(θ), so Q has a unique local maximum and one σ² works.
    Sec. 2.4: 'It is of course easy to provide counterexamples to this claim'; used to justify universal σ²≈0.1.
  • domain assumption Large networks are fairly unencumbered by local optima, and this carries over to DSML.
    Sec. 1.4: 'hopefully this will hold for DSML'; needed because the greedy algorithm can get stuck.
invented entities (1)
  • latent intelligence z no independent evidence
    purpose: Conceptual hidden quantity whose apex occurs at µ while measured performance distinguishes trajectories only at ν; motivates the two-timescale cycle.
    Defined only through the O-U toy model (dz=-z dt+dW); no measurable handle in real DSML systems.

pith-pipeline@v1.3.0-alltime-deepseek · 14868 in / 12314 out tokens · 123183 ms · 2026-07-31T23:30:16.452583+00:00 · methodology

0 comments
read the original abstract

Gradient descent scales well to large models, but becomes unstable over long time horizons. Gradient-free optimizers can scale to arbitrary timespans, but are hobbled by high dimensions. Since learning occurs in large models over long timescales, neither of these approaches is likely to produce traits which can accelerate the learning process. Instead, we propose a meta-learning algorithm in which the agent learns to modify its own weights and biases. Our algorithm consists of an inner loop, wherein the agent performs some high-dimensional optimization upon itself, and an outer loop, wherein we perform some low-dimensional optimization upon the inner loop. Since the outer loop handles very few parameters, standard zeroth-order methods may be used.

Figures

Figures reproduced from arXiv: 2607.23925 by Aria Yom.

Figure 1
Figure 1. Figure 1: The taffy map. There are essentially two possibilities. If the horizon is distant, and if the learning landscape is barren afar, then there is little need for us to wander beyond the gradient’s field of view. But if the horizon is near, and if the world beyond the horizon is rich with possibility, then to rely on the gradient alone as our compass would be to abandon the most fruitful lands to remain foreve… view at source ↗
Figure 2
Figure 2. Figure 2: First-order and zeroth-order gradient estimates at [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Time and greed. In the short-term case, 3 mutants are generated at each time step, with the best surviving and reproducing. The long-term plot is similar, but the system is simulated 3 steps into the future before crowning the winner. In a more realistic scenario, “long-term” would likely mean much more than 3 steps. Gray dashed lines denote paths sampled. Black dashed lines denote the winning paths from e… view at source ↗
Figure 4
Figure 4. Figure 4: Simulation of the system dz = −zdt + dW, dy = (−y + z 3 )dt, for a Wiener process W. Initial conditions z = 1 and y = .5 were used for 10 random trajectories. The optimal mutation timescale µ is highlighted in the upper z plot, while the optimal evaluation timescale ν is highlighted in the lower y plot. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Shapes of S and Q. The solid blue line represents the limiting case where S ∝ √ θ − √ θ0. The solid orange line represents the limiting case where S saturates to its maximum immediately upon crossing θ > θ0. The dashed line represents a “typical” case in between these two limits. The gray horizontal lines are of unit length and show that fluctuations of size σ ≪ 1 should be well-tolerated for any such conc… view at source ↗

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

72 extracted references · 5 canonical work pages

  1. [1]

    Learning to learn by gradient descent by gradient descent , year =

    Andrychowicz, Marcin and Denil, Misha and Colmenarejo, Sergio G\'. Learning to learn by gradient descent by gradient descent , year =. Proceedings of the 30th International Conference on Neural Information Processing Systems , pages =

  2. [2]

    Selection and Reinforcement Learning for Combinatorial Optimization

    Berny, A. Selection and Reinforcement Learning for Combinatorial Optimization. Parallel Problem Solving from Nature PPSN VI. 2000

    2000

  3. [3]

    Language Models are Few-Shot Learners , url =

    Brown, Tom and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared D and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel and Wu, Jeffrey and Winte...

  4. [4]

    Proceedings of the 31st International Conference on Machine Learning , pages =

    DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition , author =. Proceedings of the 31st International Conference on Machine Learning , pages =. 2014 , editor =

    2014

  5. [5]

    Siamese Neural Networks for One-Shot Image Recognition , author=

  6. [6]

    Proceedings of The 33rd International Conference on Machine Learning , pages =

    Unitary Evolution Recurrent Neural Networks , author =. Proceedings of The 33rd International Conference on Machine Learning , pages =. 2016 , editor =

    2016

  7. [7]

    2016 , eprint=

    Layer Normalization , author=. 2016 , eprint=

    2016

  8. [8]

    Bäck, Thomas H. W. and Kononova, Anna V. and van Stein, Bas and Wang, Hao and Antonov, Kirill A. and Kalkreuth, Roman T. and de Nobel, Jacob and Vermetten, Diederick and de Winter, Roy and Ye, Furong , title =. Evolutionary Computation , volume =. 2023 , month =. doi:10.1162/evco_a_00325 , url =

    doi:10.1162/evco_a_00325 2023

  9. [9]

    1992 , institution =

    Jurgen Schmidhuber , title =. 1992 , institution =

    1992

  10. [10]

    1990 , institution =

    Yoshua Bengio and Samy Bengio and Jocelyn Cloutier , title =. 1990 , institution =

    1990

  11. [11]

    and Simard, P

    Bengio, Y. and Simard, P. and Frasconi, P. , title =. 1994 , issue_date =. doi:10.1109/72.279181 , month = mar, pages =

    doi:10.1109/72.279181 1994

  12. [12]

    Neural Processing Letters , volume =

    Yoshua Bengio and Samy Bengio and Jocelyn Cloutier , title =. Neural Processing Letters , volume =. 1995 , number =

    1995

  13. [13]

    Natural Computing , volume =

    Hans-Georg Beyer and Hans-Paul Schwefel , title =. Natural Computing , volume =. 2002 , doi =

    2002

  14. [14]

    1990 , institution =

    David Chalmers , title =. 1990 , institution =

    1990

  15. [15]

    CoRR , volume =

    Sarath Chandar and Chinnadhurai Sankar and Eugene Vorontsov and Samira Ebrahimi Kahou and Yoshua Bengio , title =. CoRR , volume =. 2019 , url =. 1902.06704 , timestamp =

    Pith/arXiv arXiv 2019

  16. [16]

    , journal=

    Bharath, B and Borkar, V.S. , journal=. Robust parameter optimization of hidden Markov Models , year=

  17. [17]

    CoRR , volume =

    Jacob Devlin and Ming. CoRR , volume =. 2018 , url =. 1810.04805 , timestamp =

    Pith/arXiv arXiv 2018

  18. [18]

    Journal of Machine Learning Research , year =

    John Duchi and Elad Hazan and Yoram Singer , title =. Journal of Machine Learning Research , year =

  19. [19]

    Elsken, Thomas and Metzen, Jan Hendrik and Hutter, Frank , title =. J. Mach. Learn. Res. , month = jan, pages =. 2019 , issue_date =

    2019

  20. [20]

    Yu. M. Ermol’ev , journal=. On the method of generalized stochastic gradients and quasi-F. 1969 , volume=

    1969

  21. [21]

    Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =

    Finn, Chelsea and Abbeel, Pieter and Levine, Sergey , title =. Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =. 2017 , publisher =

    2017

  22. [22]

    Lillicrap and Mohammad Norouzi and Jimmy Ba , title =

    Danijar Hafner and Timothy P. Lillicrap and Mohammad Norouzi and Jimmy Ba , title =. CoRR , volume =. 2020 , url =. 2010.02193 , timestamp =

    Pith/arXiv arXiv 2020

  23. [23]

    Hannun and Carl Case and Jared Casper and Bryan Catanzaro and Greg Diamos and Erich Elsen and Ryan Prenger and Sanjeev Satheesh and Shubho Sengupta and Adam Coates and Andrew Y

    Awni Y. Hannun and Carl Case and Jared Casper and Bryan Catanzaro and Greg Diamos and Erich Elsen and Ryan Prenger and Sanjeev Satheesh and Shubho Sengupta and Adam Coates and Andrew Y. Ng , title =. CoRR , volume =. 2014 , url =. 1412.5567 , timestamp =

    Pith/arXiv arXiv 2014

  24. [24]

    Proceedings of the Sixth International Conference on Genetic Algorithms , pages =

    Nikolaus Hansen and Andreas Ostermeier and Andreas Gawelczyk , title =. Proceedings of the Sixth International Conference on Genetic Algorithms , pages =

  25. [25]

    and Ostermeier, A

    Hansen, N. and Ostermeier, A. , booktitle=. Adapting arbitrary normal mutation distributions in evolution strategies: the covariance matrix adaptation , year=

  26. [26]

    Completely Derandomized Self-Adaptation in Evolution Strategies , year=

    Hansen, Nikolaus and Ostermeier, Andreas , journal=. Completely Derandomized Self-Adaptation in Evolution Strategies , year=

  27. [27]

    1974 , issn =

    ALOPEX: A stochastic method for determining visual receptive fields , journal =. 1974 , issn =. doi:https://doi.org/10.1016/0042-6989(74)90024-8 , url =

    work page doi:10.1016/0042-6989(74)90024-8 1974

  28. [28]

    2016 , volume =

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , booktitle =. 2016 , volume =. doi:10.1109/CVPR.2016.90 , url =

    doi:10.1109/cvpr.2016.90 2016

  29. [29]

    Diamos and Heewoo Jun and Hassan Kianinejad and Md

    Joel Hestness and Sharan Narang and Newsha Ardalani and Gregory F. Diamos and Heewoo Jun and Hassan Kianinejad and Md. Mostofa Ali Patwary and Yang Yang and Yanqi Zhou , title =. CoRR , volume =. 2017 , url =. 1712.00409 , timestamp =

    Pith/arXiv arXiv 2017

  30. [30]

    Long Short-Term Memory , year =

    Hochreiter, Sepp and Schmidhuber, J\". Long Short-Term Memory , year =. doi:10.1162/neco.1997.9.8.1735 , journal =

    doi:10.1162/neco.1997.9.8.1735 1997

  31. [31]

    2022 , volume=

    Hospedales, Timothy and Antoniou, Antreas and Micaelli, Paul and Storkey, Amos , journal=. 2022 , volume=. doi:10.1109/TPAMI.2021.3079209 , url =

    arXiv 2022

  32. [32]

    CoRR , volume =

    Jeremy Howard and Sebastian Ruder , title =. CoRR , volume =. 2018 , url =. 1801.06146 , timestamp =

    Pith/arXiv arXiv 2018

  33. [33]

    Hubert, Thomas and Mehta, Rishi and Sartran, Laurent and Horváth, Miklós Z. and Žužić, Goran and Wieser, Eric and Huang, Aja and Schrittwieser, Julian and Schroecker, Yannick and Masoom, Hussain and Bertolli, Ottavia and Zahavy, Tom and Mandhane, Amol and Yung, Jessica and Beloshapka, Iuliya and Ibarz, Borja and Veeriah, Vivek and Yu, Lei and Nash, Oliver...

    2025

  34. [34]

    and Plaat, Aske , title =

    Huisman, Mike and van Rijn, Jan N. and Plaat, Aske , title =. 2021 , issue_date =. doi:10.1007/s10462-021-10004-4 , journal =

    work page doi:10.1007/s10462-021-10004-4 2021

  35. [35]

    1997 , issn =

    Improved stochastic optimization algorithms for adaptive optics , journal =. 1997 , issn =. doi:https://doi.org/10.1016/S0010-4655(96)00101-4 , url =

    work page doi:10.1016/s0010-4655(96)00101-4 1997

  36. [36]

    Brown and Benjamin Chess and Rewon Child and Scott Gray and Alec Radford and Jeffrey Wu and Dario Amodei , title =

    Jared Kaplan and Sam McCandlish and Tom Henighan and Tom B. Brown and Benjamin Chess and Rewon Child and Scott Gray and Alec Radford and Jeffrey Wu and Dario Amodei , title =. CoRR , volume =. 2020 , url =. 2001.08361 , timestamp =

    Pith/arXiv arXiv 2020

  37. [37]

    International Conference on Learning Representations , year=

    Adam: A Method for Stochastic Optimization , author=. International Conference on Learning Representations , year=

  38. [38]

    1978 , publisher=

    Stochastic Approximation Methods for Constrained and Unconstrained Systems , author=. 1978 , publisher=

    1978

  39. [39]

    Le and Navdeep Jaitly and Geoffrey E

    Quoc V. Le and Navdeep Jaitly and Geoffrey E. Hinton , title =. CoRR , volume =. 2015 , url =. 1504.00941 , timestamp =

    Pith/arXiv arXiv 2015

  40. [40]

    CoRR , volume =

    Ke Li and Jitendra Malik , title =. CoRR , volume =. 2016 , url =. 1606.01885 , timestamp =

    Pith/arXiv arXiv 2016

  41. [41]

    CoRR , volume =

    Ke Li and Jitendra Malik , title =. CoRR , volume =. 2017 , url =. 1703.00441 , timestamp =

    Pith/arXiv arXiv 2017

  42. [42]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Guided evolutionary strategies: augmenting random search with surrogate gradients , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    2019

  43. [43]

    CoRR , volume =

    Stephen Merity and Nitish Shirish Keskar and Richard Socher , title =. CoRR , volume =. 2017 , url =. 1708.02182 , timestamp =

    Pith/arXiv arXiv 2017

  44. [44]

    Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 , pages =

    Pascanu, Razvan and Mikolov, Tomas and Bengio, Yoshua , title =. Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 , pages =. 2013 , publisher =

    2013

  45. [45]

    Alec Radford and Jeffrey Wu and Rewon Child and David Luan and Dario Amodei and Ilya Sutskever , title =

  46. [46]

    1973 , publisher =

    Ingo Rechenberg , title =. 1973 , publisher =

    1973

  47. [47]

    A Simple Modification in CMA-ES Achieving Linear Time and Space Complexity

    Ros, Raymond and Hansen, Nikolaus. A Simple Modification in CMA-ES Achieving Linear Time and Space Complexity. Parallel Problem Solving from Nature -- PPSN X. 2008

    2008

  48. [48]

    and Jonsson, M.T

    Runarsson, T.P. and Jonsson, M.T. , booktitle=. Evolution and design of distributed learning rules , year=

  49. [49]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Are Emergent Abilities of Large Language Models a Mirage? , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  50. [50]

    Evolutionary Principles in Self-Referential Learning

    Jurgen Schmidhuber. Evolutionary Principles in Self-Referential Learning. 1987

    1987

  51. [51]

    , booktitle=

    Schmidhuber, J. , booktitle=. A neural network that embeds its own meta-levels , year=

  52. [52]

    Schmidhuber, J and Zhao, Jieyu and Wiering, Marco , year =

  53. [53]

    Mastering Atari, Go, chess and shogi by planning with a learned model , year=

    Schrittwieser, Julian and Antonoglou, Ioannis and Hubert, Thomas and Simonyan, Karen and Sifre, Laurent and Schmitt, Simon and Guez, Arthur and Lockhart, Edward and Hassabis, Demis and Graepel, Thore and Lillicrap, Timothy and Silver, David , journal=. Mastering Atari, Go, chess and shogi by planning with a learned model , year=

  54. [54]

    Hans-Paul Schwefel , title =

  55. [55]

    2010 , note =

    Parameter-exploring policy gradients , journal =. 2010 , note =. doi:https://doi.org/10.1016/j.neunet.2009.12.004 , url =

    doi:10.1016/j.neunet.2009.12.004 2010

  56. [56]

    Mastering the game of Go without human knowledge , year=

    Silver, David and Schrittwieser, Julian and Simonyan, Karen and Antonoglou, Ioannis and Huang, Aja and Guez, Arthur and Hubert, Thomas and Baker, Lucas and Lai, Matthew and Bolton, Adrian and Chen, Yutian and Lillicrap, Timothy and Hui, Fan and Sifre, Laurent and van den Driessche, George and Graepel, Thore and Hassabis, Demis , journal=. Mastering the ga...

  57. [57]

    , journal=

    Spall, J.C. , journal=. Review Of Stochastic Approximation Algorithms And Applications [Book Reviews] , year=

  58. [58]

    Highway Networks , journal =

    Rupesh Kumar Srivastava and Klaus Greff and J. Highway Networks , journal =. 2015 , url =. 1505.00387 , timestamp =

    Pith/arXiv arXiv 2015

  59. [59]

    , title =

    Sutton, Richard S. , title =. 2019 , month = mar, day =

    2019

  60. [60]

    Learning to Learn: Introduction and Overview

    Thrun, Sebastian and Pratt, Lorien. Learning to Learn: Introduction and Overview. Learning to Learn. 1998. doi:10.1007/978-1-4615-5529-2_1

    work page doi:10.1007/978-1-4615-5529-2_1 1998

  61. [61]

    Unnikrishnan, K. P. and Venugopal, K. P. , title =. Neural Computation , volume =. 1994 , month =. doi:10.1162/neco.1994.6.3.469 , url =

    work page doi:10.1162/neco.1994.6.3.469 1994

  62. [62]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. CoRR , volume =. 2017 , url =. 1706.03762 , timestamp =

    Pith/arXiv arXiv 2017

  63. [63]

    Wilber and Serge J

    Andreas Veit and Michael J. Wilber and Serge J. Belongie , title =. CoRR , volume =. 2016 , url =. 1605.06431 , timestamp =

    Pith/arXiv arXiv 2016

  64. [64]

    Advances and Challenges in Meta-Learning: A Technical Review , year =

    Vettoruzzo, Anna and Bouguelia, Mohamed-Rafik and Vanschoren, Joaquin and R\". Advances and Challenges in Meta-Learning: A Technical Review , year =. doi:10.1109/TPAMI.2024.3357847 , journal =

    arXiv 2024

  65. [65]

    2016 , eprint=

    Bayesian Optimization in a Billion Dimensions via Random Embeddings , author=. 2016 , eprint=

    2016

  66. [66]

    Transactions on Machine Learning Research , issn=

    Emergent Abilities of Large Language Models , author=. Transactions on Machine Learning Research , issn=. 2022 , url=

    2022

  67. [67]

    Journal of Machine Learning Research , year =

    Daan Wierstra and Tom Schaul and Tobias Glasmachers and Yi Sun and Jan Peters and Jurgen Schmidhuber , title =. Journal of Machine Learning Research , year =

  68. [68]

    Williams , title =

    R.J. Williams , title =. Mach Learn , year =

  69. [69]

    CoRR , volume =

    Jason Yosinski and Jeff Clune and Yoshua Bengio and Hod Lipson , title =. CoRR , volume =. 2014 , url =. 1411.1792 , timestamp =

    Pith/arXiv arXiv 2014

  70. [70]

    CoRR , volume =

    Wojciech Zaremba and Ilya Sutskever and Oriol Vinyals , title =. CoRR , volume =. 2014 , url =. 1409.2329 , timestamp =

    Pith/arXiv arXiv 2014

  71. [71]

    2021 , eprint=

    A New One-Point Residual-Feedback Oracle For Black-Box Learning and Control , author=. 2021 , eprint=

    2021

  72. [72]

    , journal=

    Zhang, Yan and Zhou, Yi and Ji, Kaiyi and Shen, Yi and Zavlanos, Michael M. , journal=. Boosting One-Point Derivative-Free Online Optimization via Residual Feedback , year=