Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Heuristic Learning for Active Flow Control Using Coding Agents

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Coding agents can discover explicit flow controllers that match or beat deep reinforcement learning on most standard benchmarks while remaining human-readable.

desk verdict Equal-budget coding-agent search finds inspectable AFC controllers that match or beat strongest DRL on 10/13 public benchmarks; the result is real, the main caveat is protocol/sandbox reproducibility. read the letter →

arxiv 2607.11565 v1 pith:INXHFQCS submitted 2026-07-13 cs.LG cs.AIphysics.flu-dyn

Paul Garnier , Jonathan Viquerat , Elie Hachem This is my paper
classification cs.LGcs.AIphysics.flu-dyn
keywords activeflowcontrolfluiddynamicsheuristiclearningcodingagentsreinforcementbenchmarksinterpretable
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

Active flow control is hard because the dynamics are nonlinear, sensors see only part of the state, and every trial costs an expensive fluid simulation. Deep reinforcement learning has become a common attack on that problem, but it usually needs many rollouts and returns neural policies whose decisions are hard to inspect. This paper tests a different search: modern coding agents iteratively write, evaluate, and revise short executable feedback laws, using only the public benchmark interface and exactly the same total simulation budget as the DRL baselines. On thirteen one-, two-, and three-dimensional control environments the best heuristics match or outperform the strongest available DRL policy in ten cases, and the surviving programs are compact enough to read, ablate, and transfer. The stake for a sympathetic reader is a practical complementary route to controllers that stay competitive while restoring physical interpretability.

What carries the argument

The constrained heuristic-learning protocol: a coding agent proposes executable controller programs, evaluates them only through the public benchmark interface, revises them from a public ledger of rollout evidence, filters invalid policies, and charges the full simulation budget even for early-rejected candidates.

What would settle it

Re-run the full 13-environment suite inside a sealed sandbox that blocks all non-public API access and enforces identical total environment steps; if the best valid heuristics then fall below the strongest DRL baseline on a clear majority of tasks, the central claim fails.

Watch

Extended reading notes

Core claim

Under identical public environments, observations, actions, rewards, and total simulation budgets, a constrained coding-agent search over executable controller source code discovers explicit feedback laws that match or outperform the strongest available DRL baseline among PPO, SAC, and TD3 in 10 of 13 active flow-control benchmarks, while remaining compact and directly inspectable.

Load-bearing premise

The comparison is fair only if the agent truly never peeks at hidden simulator state or rewrites the reward, and if every early-stopped trial is fully charged against the same simulation budget as deep reinforcement learning.

Editorial extensions

If this is right

  • Explicit controllers can compete with neural DRL policies on standard flow-control scores without giving up inspectability.
  • Discovered laws often reduce to physically meaningful mechanisms such as delayed shape feedback or local wall-velocity feedback that can be read and modified by hand.
  • Controllers can transfer to harder regimes (higher Reynolds or Rayleigh number, more actuators) by editing a few explicit constants rather than retraining a network.
  • Early search can avoid the long cold-start phase typical of DRL while still respecting the same interaction budget.
  • A fixed harness with identical budgets turns multi-agent comparison into a controlled test of coding-agent capability itself.

Reading between the lines

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

  • The same public-interface protocol could transfer to other expensive partial-observation control domains where engineers value readable laws over opaque weights.
  • As coding agents improve, the limiting factor may become validity harnesses and budget accounting rather than raw code-writing skill.
  • Hybrid pipelines that seed DRL with heuristic structure, or freeze heuristics as transparent baselines, are a natural engineering follow-on.
  • Independent sealed-sandbox replications would be the cleanest next measurement of whether interface violations can still occur under the stated protocol.
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 / 6 minor

Summary. The paper proposes a constrained heuristic-learning protocol in which coding agents iteratively write, evaluate, and revise explicit executable feedback controllers for active flow control, interacting only through public benchmark interfaces under the same total simulation budgets as DRL. On 13 BEACON and FluidGym environments spanning 1D–3D problems, the best discovered heuristics match or exceed the strongest reproduced DRL baseline (among PPO, SAC, TD3) in 10 of 13 cases (Abstract; §6; Fig. 3), while remaining compact and inspectable. Additional experiments address sample-efficiency trajectories (Fig. 6), multi-agent comparison (Codex/Claude/Fable/Gemini; Fig. 7), transfer across Re/Ra and actuator counts (§6.2.1; Fig. 8), richer vs sparse observations (§6.2.2–6.2.3; Fig. 9), and policy anatomy with ablations (Shkadov delayed shape feedback; TCF local wall feedback; §7; Algorithms 2–13). Prompts and code are released.

Significance. If the equal-budget, public-interface comparisons hold under independent audit, the work is a credible complementary alternative to DRL for flow control: competitive performance with human-readable laws, physically interpretable mechanisms (e.g., transport delay and wave-shape features in Shkadov; normalized local wall feedback in TCF), and practical transfer by editing a few constants rather than retraining networks. Strengths include standardized environments with reproduced DRL baselines, explicit budget accounting including early-rejected rollouts, controller ablations, long-horizon cylinder validation (Fig. 5), and open prompts/source. That combination is valuable for both control practice and reproducible agent-based scientific discovery.

major comments (3)
  1. [§4.1 Explicit constraints; Algorithm 1] §4.1 and Algorithm 1: Validity of the central equal-budget claim rests on prompt-level constraints, static source checks, and a public ledger rather than a hard technical sandbox that independent auditors can re-run to prove no hidden-state access, reward edits, or environment cloning. The paper correctly charges early-rejected steps to the budget, but the manuscript should state more precisely what was logged for every rejected/invalid candidate (hashes of controller source, interface call traces, and whether any agent attempt was discarded post hoc), and ideally provide a minimal machine-checkable validity harness or replay package so the 10/13 result cannot be challenged as harness-dependent.
  2. [§6 Main Results; Fig. 3; Fig. 6] §6 and Fig. 3: The headline “10 of 13” comparison is reported as point scores of the best heuristic vs the best DRL algorithm per environment. Fig. 6 and the text assert that repeated heuristic runs converge to similar controllers, but the main aggregate figure does not report multi-seed means/variance for the coding-agent search (or confidence intervals on the normalized score). Because free parameters include agent model, reasoning effort, prompts, and early-rejection thresholds, a short multi-seed table (or error bars) for at least the BEACON suite and the three underperforming environments is needed to make the 10/13 claim statistically load-bearing rather than best-of-search.
  3. [§6; Fig. 3] §6 (environments where heuristic < best DRL) and §7: The paper emphasizes wins and interpretability but gives limited systematic failure analysis for the three environments where heuristics underperform. A brief diagnosis—whether the search exhausted the budget on invalid/weak candidates, whether the optimal policy is hard to express as a compact program, or whether DRL exploits non-interpretable correlations—would strengthen the claim that heuristic learning is a complementary alternative rather than only reporting the favorable aggregate.
minor comments (6)
  1. [Fig. 3; §6] Fig. 3 caption and §6: Clarify the exact normalization (improvement from no-control or from “initial DRL control level”) and whether “match” means numerical equality within a stated tolerance or within reported DRL variance.
  2. [§5 Experimental Setup] §5: Agent names and model versions (GPT-5.5/Codex, Claude Opus 4.7, Fable 5, Gemini 3 Pro) and “xhigh” reasoning settings will age quickly; pin exact API model IDs and dates used for the reported runs.
  3. [Appendix B] Appendix B Algorithms 2–13: Several numerical coefficients are given to many decimals without stating whether they were frozen after search or re-tuned on validation seeds; a one-line note on selection criterion (best seed, mean over seeds, etc.) would help reproducibility.
  4. [Fig. 6] Fig. 6: Extending the last heuristic value when diagnostics used fewer steps is reasonable but should be marked (e.g., dashed continuation) so readers do not misread flat late trajectories as continued search.
  5. [Throughout; Fig. 2; Table 2] Typos/formatting: title line breaks (“HEURISTICLEARNING”, “CODINGAGENTS”); “Flo w”, “Ra yleigh”, “V or t e x” in Fig. 2; “V ortex-induced” in Table 2; ensure consistent hyphenation of Rayleigh–Bénard.
  6. [§2.3] §2.3: The program-synthesis related-work survey is useful; a short explicit contrast with classical gain-scheduled/PID tuning and symbolic regression for control would situate the contribution for the fluid-dynamics audience.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: equal-budget heuristic vs DRL scores are external empirical outcomes, not forced by definition or self-citation.

full rationale

The paper’s load-bearing claim is empirical: under fixed public interfaces and identical simulation budgets, coding-agent search yields explicit controllers that match or beat the strongest DRL baseline in 10 of 13 environments (Abstract; §6; Fig. 3). Performance is defined by external simulators and public rewards (BEACON, FluidGym); DRL baselines (PPO, SAC, TD3) are independent methods on the same interfaces; controller gains, delays, and filters are search outputs, not inputs that force the headline comparison by construction. Related-work self-citations (e.g. authors’ BEACON suite and prior DRL reviews) supply benchmarks and context but do not justify uniqueness theorems or smuggle an ansatz that makes the 10/13 result tautological. Policy anatomy (§7) interprets discovered structure after evaluation; transfer and sparsity experiments re-evaluate on held-out settings rather than renaming fitted quantities as predictions. Protocol validity rests on harness constraints, which is a reproducibility caveat, not circular derivation. Score 0 with empty steps is therefore appropriate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim is empirical and rests on standard MDP control formulation, public benchmark interfaces, equal interaction budgets, and a prompt-constrained coding-agent loop. Free parameters are mostly protocol and model choices plus the many numerical gains the search itself produces (outputs, not inputs to the claim). No new physical particles or forces are postulated; the main invented construct is the constrained heuristic-learning protocol and public search ledger.

free parameters (4)
  • Per-controller gains, delays, smoothers, and feature weights (e.g., Shkadov D=11, g≈0.3475, α=0.84; TCF α≈0.264, β_u≈0.2
    These numbers are fitted by the coding-agent search to maximize benchmark score; they are outputs of the method, not fixed physical constants, and the performance claim depends on them.
  • Coding-agent model and reasoning settings (Codex/GPT-5.5 xhigh; Claude Opus 4.7; Fable 5; Gemini 3 Pro)
    Headline results use specific commercial models and effort flags; different agents yield different normalized averages (Codex 1.13 vs Gemini 0.50 on BEACON).
  • Prompt suite and invalid-policy filter rules
    The search behavior is steered by author-written prompts and validity checks; these are free design choices of the protocol, not derived from first principles.
  • Early-rejection diagnostics thresholds for weak candidates
    When evaluations are stopped early, steps are still charged, but the decision to stop is a heuristic that affects how many candidates are tried within budget.
assumptions (4)
  • domain assumption Flow-control tasks are well modeled as MDPs with the given public observations, actions, rewards, and horizons (standard RL formulation).
    §3 adopts the usual sequential decision setup without proving that the chosen sensors/rewards are optimal for the engineering objective.
  • ad hoc to paper Equal total environment interaction counts make DRL and coding-agent search comparable even though search structures differ (many short programs vs one trained network).
    §4.1 Search budget accounting defines fairness this way; it is a protocol choice, not a theorem that sample efficiency is measured identically.
  • ad hoc to paper Prompt-level constraints plus static validity checks suffice to keep agents on the public interface (no hidden state, no reward edits, no pretrained NN policies).
    §4.1 Prompt-induced protocol violations and Validity constraints; enforcement is harness/process, not a formal sandbox proof.
  • domain assumption Strongest DRL baselines among PPO/SAC/TD3 on the official open-source stacks are adequate reference points for each environment.
    §5–6; common in the field but leaves open stronger or differently tuned RL methods.
invented entities (2)
  • Constrained heuristic-learning protocol (public ledger L, invalid-policy filtering, equal-budget coding-agent loop) independent evidence
    purpose: Define a reproducible search procedure that produces only explicit train-free controllers comparable to DRL.
    Introduced in §4 and Algorithm 1 as the paper’s methodological object; independent evidence is the released prompts/code and multi-environment scores, not an external physical measurement.
  • Public search ledger of scores, traces, ablations, and failed hypotheses independent evidence
    purpose: Accumulate textual/experimental memory that guides subsequent code edits without gradient updates.
    Described in §4 as the non-gradient optimization state; exists as an experimental artifact of the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Heuristic Learning for Active Flow Control Using Coding Agents." pith.science (2026). https://pith.science/paper/INXHFQCS

@misc{pith2026260711565,
  author       = {Pith},
  title        = {Pith review of: Heuristic Learning for Active Flow Control Using Coding Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INXHFQCS}},
  note         = {Machine review of arXiv:2607.11565}
}
read the original abstract

Active flow control involves nonlinear dynamics, partial observations, and computationally expensive simulations, making controller design particularly challenging. Deep reinforcement learning (DRL) has emerged as a powerful framework for such problems, but its success typically relies on large numbers of simulator interactions and produces neural-network policies whose decision process often remains difficult to interpret. In this work, we investigate a different paradigm: instead of optimizing neural-network parameters, we use modern coding agents to search directly for explicit executable feedback laws. We introduce a constrained heuristic-learning protocol in which an agent iteratively proposes, evaluates, and revises controller implementations while interacting exclusively through the public benchmark interface. The proposed framework is evaluated on 13 active flow-control benchmarks spanning one, two, and three-dimensional problems and compared against the strongest available DRL baselines under identical simulation budgets. The discovered heuristic controllers match or outperform the best DRL policy in 10 of the 13 environments while remaining compact, interpretable, and directly inspectable. Beyond aggregate performance, the resulting controllers reveal physically meaningful feedback mechanisms, transfer successfully across more challenging configurations, and remain competitive under varying Reynolds and Rayleigh numbers, actuator counts, and observation sparsity. These results suggest that heuristic learning through coding agents constitutes a credible and complementary alternative to conventional reinforcement learning, combining competitive performance with physically interpretable controller representations. Prompts and source code are available at https://github.com/DonsetPG/fluid-heuristic-learning.

Figures

Figures reproduced from arXiv: 2607.11565 by the authors.

Figure 1
Figure 1. Comparison between the standard reinforcement-learning interaction loop and heuristic learning, where the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Presentation of the different environments. Abstract schematic [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Aggregate comparison between the best coding-agent heuristic and the best DRL baseline on each environment. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Drag reduction obtained by the heuristic controllers in the turbulent-channel and cylinder-control cases. We [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Long-horizon 400-step validation drag on [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Score evolution as a function of consumed environment steps. DRL curves are shown together with the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Agent comparison on the BEACON environments. Each bar reports the normalized score reached by the [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Shkadov actuator transfer from 5 to 10 jets. The left panel compares the 5-jet and 10-jet PPO references with [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Sparse-observation control. The left panel shows the sensor layouts and the bottom actuators; the right panels [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Same-state comparison between PPO and the heuristic on [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 5 canonical work pages

  1. [1]

    Vincent Belus, Jean Rabault, Jonathan Viquerat, Zhizhao Che, Elie Hachem, and Ulysse Reglade

    doi: 10.1017/jfm.2019.62. Vincent Belus, Jean Rabault, Jonathan Viquerat, Zhizhao Che, Elie Hachem, and Ulysse Reglade. Exploiting locality and translational invariance to design effective deep reinforcement learning control of the 1-dimensional unstable falling liquid film.AIP Advances, 9(12):125014,

    doi:10.1017/jfm.2019.62 2019

  2. [2]

    Gerben Beintema, Alessandro Corbetta, Luca Biferale, and Federico Toschi

    doi: 10.1063/1.5132378. Gerben Beintema, Alessandro Corbetta, Luca Biferale, and Federico Toschi. Controlling rayleigh–bénard convection via reinforcement learning.Journal of Turbulence, 21(9–10):585–605,

    work page doi:10.1063/1.5132378

  3. [3]

    Elie Hachem, H

    doi: 10.1080/14685248.2020.1797059. Elie Hachem, H. Ghraieb, Jonathan Viquerat, Aurélien Larcher, and Philippe Meliga. Deep reinforcement learning for the control of conjugate heat transfer.Journal of Computational Physics, 436:110317,

    doi:10.1080/14685248.2020.1797059 2020

  4. [4]

    doi: 10.1016/j.jcp.2021. 110317. Guido Novati, Siddhartha Verma, Dmitry Alexeev, Diego Rossinelli, Wim M. van Rees, and Petros Koumoutsakos. Synchronisation through learning for two self-propelled swimmers.Bioinspiration & Biomimetics, 12(3):036001,

    doi:10.1016/j.jcp.2021 2021

  5. [5]

    Siddhartha Verma, Guido Novati, and Petros Koumoutsakos

    doi: 10.1088/1748-3190/aa6311. Siddhartha Verma, Guido Novati, and Petros Koumoutsakos. Efficient collective swimming by harnessing vortices through deep reinforcement learning.Proceedings of the National Academy of Sciences, 115(23):5849–5854,

    work page doi:10.1088/1748-3190/aa6311

  6. [6]

    Jonathan Viquerat, Jean Rabault, Alexander Kuhnle, Hassan Ghraieb, Aurélien Larcher, and Elie Hachem

    doi: 10.1073/pnas.1800923115. Jonathan Viquerat, Jean Rabault, Alexander Kuhnle, Hassan Ghraieb, Aurélien Larcher, and Elie Hachem. Direct shape optimization through deep reinforcement learning.Journal of Computational Physics, 428:110080,

    doi:10.1073/pnas.1800923115

  7. [7]

    Dixia Fan, Liu Yang, Zhicheng Wang, Michael S

    doi: 10.1016/j.jcp.2020.110080. Dixia Fan, Liu Yang, Zhicheng Wang, Michael S. Triantafyllou, and George Em Karniadakis. Reinforcement learning for bluff body active flow control in experiments and simulations.Proceedings of the National Academy of Sciences, 117(42):26091–26098,

    doi:10.1016/j.jcp.2020.110080 2020

  8. [8]

    Jean Rabault and Alexander Kuhnle

    doi: 10.1073/pnas.2004939117. Jean Rabault and Alexander Kuhnle. Accelerating deep reinforcement learning strategies of flow control through a multi-environment approach.Physics of Fluids, 31(9):094105,

    doi:10.1073/pnas.2004939117

Show all 35 references
  1. [9]

    Jonathan Viquerat and Elie Hachem

    doi: 10.1063/1.5116415. Jonathan Viquerat and Elie Hachem. Parallel bootstrap-based on-policy deep reinforcement learning for continuous fluid flow control applications.Fluids, 8(7):208,

    doi

  2. [10]

    doi: 10.3390/fluids8070208. Marcin Andrychowicz, Anton Raichuk, Piotr Sta ´nczyk, Manu Orsini, Sertan Girgin, Raphael Marinier, Léonard Hussenot, Matthieu Geist, Olivier Pietquin, Marcin Michalski, Sylvain Gelly, and Olivier Bachem. What matters in on-policy reinforcement lear...

    doi

  3. [11]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio

    doi: 10.1063/5.0128446. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. InInternational Conference on Learning Representations,

    doi

  4. [12]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    16 HEURISTICLEARNING FORFLUIDDYNAMICS- JULY14, 2026 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems, volume 30,

    2026

  5. [14]

    Carlos E

    URLhttps://arxiv.org/abs/2107.03374. Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R. Narasimhan. SWE-bench: Can language models resolve real-world github issues? InThe Twelfth International Conference on Learning Representations,

    arXiv

  6. [15]

    Anthropic

    URLhttps://openreview.net/forum?id=VTF8yNQM66. Anthropic. Claude code by anthropic. https://www.anthropic.com/product/claude-code, 2026a. Accessed: 2026-05-19. OpenAI. Introducing codex. https://openai.com/index/introducing-codex/, May

    2026

  7. [16]

    Introducing the codex app

    OpenAI. Introducing the codex app. https://openai.com/index/introducing-the-codex-app/ , February 2026a. Published February 2, 2026; updated March 4,

    2026

  8. [17]

    Feng Ren, Jean Rabault, and Hui Tang

    doi: 10.1063/5.0006492. Feng Ren, Jean Rabault, and Hui Tang. Applying deep reinforcement learning to active flow control in weakly turbulent conditions.Physics of Fluids, 33(3):037121,

    doi

  9. [18]

    Paul Garnier, Jonathan Viquerat, Jean Rabault, Aurélien Larcher, Alexander Kuhnle, and Elie Hachem

    doi: 10.1063/5.0037371. Paul Garnier, Jonathan Viquerat, Jean Rabault, Aurélien Larcher, Alexander Kuhnle, and Elie Hachem. A review on deep reinforcement learning for fluid mechanics.Computers & Fluids, 225:104973,

    doi

  10. [19]

    2021.104973

    doi: 10.1016/j.compfluid. 2021.104973. Elie Hachem, Aditya Vishwasrao, M. Renault, Jonathan Viquerat, and Philippe Meliga. Reinforcement learning for cooling rate control during quenching.International Journal of Numerical Methods for Heat & Fluid Flow, 34(8),

    2021 doi

  11. [20]

    Emanuel Todorov, Tom Erez, and Yuval Tassa

    doi: 10.1613/jair.3912. Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033,

    doi

  12. [21]

    doi: 10.1109/IROS.2012. 6386109. Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym,

    2012 doi

  13. [22]

    Jannis Becktepe, Aleksandra Franz, Nils Thuerey, and Sebastian Peitz

    doi: 10.3390/app14093561. Jannis Becktepe, Aleksandra Franz, Nils Thuerey, and Sebastian Peitz. Plug-and-play benchmarking of reinforcement learning algorithms for large-scale flow control,

    doi

  14. [23]

    Yoonho Lee, Joseph Boen, and Chelsea Finn

    URLhttps://arxiv.org/abs/2601.15015. Yoonho Lee, Joseph Boen, and Chelsea Finn. Feedback descent: Open-ended text optimization via pairwise comparison. InarXiv preprint arXiv:2511.07919,

    arXiv

  15. [24]

    Gepa: Reflective prompt evolution can outperform reinforcement learning.arXiv preprint arXiv:2507.19457,

    Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, et al. Gepa: Reflective prompt evolution can outperform reinforcement learning.arXiv preprint arXiv:2507.19457,

    arXiv

  16. [25]

    gradient descent

    Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng. Automatic prompt optimization with “gradient descent” and beam search.arXiv preprint arXiv:2305.03495,

    arXiv

  17. [26]

    Self-refine: Iterative refinement with self-feedback.Advances in neural information processing systems, 36:46534–46594,

    17 HEURISTICLEARNING FORFLUIDDYNAMICS- JULY14, 2026 Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback.Advances in neural info...

    2026

  18. [27]

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen

    URLhttps://arxiv.org/abs/2406.07496. Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. InThe Twelfth International Conference on Learning Representations,

    arXiv

  19. [28]

    GitHub repository

    URL https://github.com/ algorithmicsuperintelligence/openevolve. GitHub repository. Alexander Novikov, Ngân V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco JR Ruiz, Abbas Mehrabian, et al. Alphaevolve...

    arXiv

  20. [29]

    URLhttps://arxiv.org/abs/2206.08896. Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with la...

    arXiv

  21. [30]

    URLhttps://openreview.net/forum?id=t9U3LW7JVX. Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. Aflow: Automating agentic workflow generation,...

    arXiv

  22. [31]

    URLhttps://arxiv.org/abs/2211.15661. OpenAI. Introducing gpt-5.5. https://openai.com/index/introducing-gpt-5-5/ , April 2026b. Published April 23,

    arXiv

  23. [32]

    Using gpt-5.5

    OpenAI. Using gpt-5.5. https://developers.openai.com/api/docs/guides/latest-model, 2026c. Ac- cessed: 2026-05-19. OpenAI. Codex models.https://developers.openai.com/codex/models, 2026d. Accessed: 2026-05-19. Anthropic. Introducing claude opus 4.7. https://www.anthropic.com/new...

    2026

  24. [33]

    Model configuration

    Anthropic. Model configuration. https://code.claude.com/docs/en/model-config, 2026d. Claude Code documentation; accessed: 2026-05-19. Anthropic. Claude fable 5 and claude mythos

    2026

  25. [34]

    Accessed: 2026-06-10

    https://www.anthropic.com/news/ claude-fable-5-mythos-5, 2026e. Accessed: 2026-06-10. 18 HEURISTICLEARNING FORFLUIDDYNAMICS- JULY14, 2026 Google. A new era of intelligence with gemini

    2026

  26. [35]

    Codex cli.https://developers.openai.com/codex/cli, 2026e

    OpenAI. Codex cli.https://developers.openai.com/codex/cli, 2026e. Accessed: 2026-05-19. Google Cloud. Gemini cli. https://docs.cloud.google.com/gemini/docs/codeassist/gemini-cli,

    2026

  27. [36]

    A Environment Suite The benchmark suite contains 13 environments: 6 larger FluidGym flow-control cases and 7 compact BEACON control cases

    Accessed: 2026-05-19. A Environment Suite The benchmark suite contains 13 environments: 6 larger FluidGym flow-control cases and 7 compact BEACON control cases. Table 2 reports the task definition, principal physical parameters, default episode horizon, and action-step clock u...

    2026

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.