Pith. sign in

REVIEW 4 major objections 5 minor 82 references

The paper argues that LLM program optimizers do not need to reason over performance scores: generating diverse atomic edits blind to scores and then selecting the best-scoring subset avoids the weakest-link effect and outperforms first-orde

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

2026-08-03 16:40 UTC pith:TSBDFYXP

load-bearing objection HERO is a clean, well-tested idea—recombine atomic edits instead of accepting the whole bundle—but the headline efficiency claim is not yet proven because total evaluator cost is never accounted for. the 4 major comments →

arxiv 2607.28947 v1 pith:TSBDFYXP submitted 2026-07-31 cs.LG

Overcoming the Weakest-Link Effect in LLM-Driven Program Optimization via Heterogeneous Edit Recombination

Jingwen Fu , Zhen Liu , Yuhan Liu , He Zhang , Nanning Zheng This is my paper
classification cs.LG
keywords LLM program optimizationzeroth-order optimizationweakest-link effectatomic edit recombinationcredit assignmentprogram searchsubset selectionevaluator-guided search
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 tries to establish that large language models can optimize programs without reading their own performance scores. Existing optimizers treat past scores as a textual gradient, but credit assignment from a scalar score back to program parts becomes unreliable when the program-to-score mapping is complex. The proposed method, HERO, asks the model only for many small, non-overlapping candidate edits, then uses the evaluator as a black-box comparator to keep the subset of edits that improves the score. Across combinatorial construction, strategy games, agent system design, and path planning, this zeroth-order strategy reaches better solutions, converges faster, and spends fewer tokens than first-order baselines, with the largest gaps appearing exactly where score-to-performance reasoning is hardest.

Core claim

HERO's central claim is that the bottleneck in LLM-driven program optimization is not generating directions but composing them: when a batch of edits is accepted or rejected as a whole, a single harmful edit vetoes all useful ones, producing the weakest-link effect. The paper formalizes a zeroth-order alternative: sample heterogeneous, pairwise non-overlapping atomic edits from the model's prior, withholding all scores, then choose the subset of edits whose recombined program has the highest measured score. This turns optimization into a ranking problem over subsets rather than an inverse problem over scores. Empirically, on a hard strategy-game setting HERO reaches 0.85 win-rate while first

What carries the argument

Subset recombination of edits under a commutativity constraint. The LLM proposes SEARCH/REPLACE edits whose regions do not overlap, so compositions commute and every subset S defines a valid recombined program A_S(x). Selection solves argmax_S [f(A_S(x))-f(x)]/||S|| using evaluator scores, which under the paper's additive approximation f(A_{1:n}(x))-f(x) ≈ Σ δ_i reduces to keeping exactly the edits with positive marginal gains. This mechanism converts the LLM's noisy partial directions into a reliably aligned ascent direction without ever feeding scores back to the model.

Load-bearing premise

The method stands on the assumption that edits touching disjoint code regions change the final score nearly independently, so the best subset of edits is essentially the set of individually beneficial edits; if non-overlapping edits interact strongly in the objective, subset selection can keep a combination that hurts.

What would settle it

Measure pairwise interactions directly: for a task, compute the marginal gain δ_i of each edit alone and the joint gain δ_ij of each pair. If many pairs satisfy δ_i>0, δ_j>0 but δ_ij<0, or δ_i+δ_j>0 while the pair's actual gain is negative, the additivity assumption fails and HERO's argmax subset rule would systematically retain harmful pairs, making full-bundle acceptance or an interaction-aware selector better.

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

If this is right

  • If HERO is right, LLM optimizers can be decoupled from score-reasoning and still make progress, lowering token costs because prompts omit scored examples and reasoning chains.
  • The hardest-to-reason-about tasks, not the easiest, become the sweet spot: the method's advantage widens as the program-to-score mapping grows more entangled.
  • A single failed edit no longer cascades: recombination lets each step compound beneficial changes across steps, which explains the observed faster convergence.
  • In agent-system design, improvements found with a small model transfer to a larger model and to held-out instances, suggesting selected edits encode structural properties rather than overfitting.
  • A finite evaluation history cannot uniquely determine the correct update direction under an unconstrained score-to-performance map, so first-order methods face an information-theoretic ceiling in complex regimes.

Where Pith is reading between the lines

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

  • The paper's ranking-based selection suggests HERO would tolerate noisy or approximate evaluators as long as they preserve subset ordering, which could extend the method to settings where exact evaluation is expensive or stochastic.
  • One could hybridize: in low-prior regimes where the model has less latent knowledge, seed HERO with a first-order signal and let recombination refine it; the paper leaves this combination open.
  • The commutativity and non-overlap requirements limit edits to independent regions; a dependency-aware variant that groups or orders interacting edits could extend the mechanism to tightly coupled programs, at the cost of a harder subset search.
  • If the additive approximation holds only approximately, selection still tends to favor subsets with many independently positive edits, so the method may degrade gracefully rather than fail sharply.

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

4 major / 5 minor

Summary. The paper introduces HERO, a zeroth-order LLM-based program optimizer that separates edit generation from edit selection. At each step, an LLM proposes multiple heterogeneous, non-overlapping atomic edits without seeing performance scores; a selection phase then evaluates recombined subsets with a scalar evaluator and keeps the best-scoring subset. The paper argues that this avoids the 'weakest-link effect' of all-or-nothing batch acceptance, where one harmful edit negates many beneficial ones. The theoretical motivation includes a latent-objective model of LLM proposals, an additivity approximation for non-overlapping edits, and an impossibility result showing that finite evaluation histories cannot uniquely determine first-order update directions. Experiments cover circle packing, Othello and Battleship strategy games, LLM agent system design on AIME, and grid-based path planning, reporting faster convergence, fewer tokens, and higher final scores than evolutionary and textual-gradient baselines.

Significance. If the empirical claims hold, the paper makes a substantive conceptual contribution: it challenges the prevailing assumption that LLM optimizers must reason over performance scores, and shows that a recombination-only optimizer can outperform first-order methods on hard program-score mappings. The multi-domain evaluation, cross-model consistency, and ablations directly targeting the recombination mechanism are genuine strengths. The formal impossibility result in Appendix C is self-contained, and the authors' explicit correction of the earlier probabilistic statement is commendable. The main limitations are that the theoretical mechanism relies on an additivity assumption that is not justified by non-overlap alone, and that the reported efficiency gains are not controlled for evaluator cost. Both issues are load-bearing for the paper's central claims, but they are addressable within the manuscript's scope.

major comments (4)
  1. [§2.1, Eq. (4)] The additivity approximation f(α_{1:n}(x_t)) − f(x_t) ≈ Σ_i δ_i is asserted to follow from non-overlapping edits, but non-overlapping code regions do not imply additive effects on a global objective. In circle packing, moving one row or regrowing one radius changes the feasible configuration for other circles; in Othello, changing one heuristic alters the entire game trajectory. When additivity fails, the claimed equivalence between Eq. (7)'s argmax and retaining exactly {i : δ_i > 0} breaks, and the theoretical mechanism for neutralizing the weakest-link effect is weakened. Please either state conditions under which Eq. (4) holds, verify additivity empirically on the actual tasks by measuring individual and joint effects, or reframe the additivity claim as a heuristic motivation rather than a formal derivation.
  2. [Abstract and §3 (efficiency claims)] The central efficiency claim—'converges substantially faster ... while consuming fewer tokens'—is supported only by step counts and LLM-token counts. The selection phase in Eq. (7)/(8) evaluates many edit subsets per step, and the faithful Othello evaluator runs 1,000 games per candidate. The paper never reports the number of evaluator calls per step, total simulator cost, or wall-clock time for HERO versus baselines. If first-order baselines evaluate one candidate per step and HERO evaluates dozens, the LLM-token savings may be offset by evaluator cost, making 'fewer tokens' a cost-shifting effect rather than a true efficiency gain. Please provide a per-domain accounting of evaluator calls, total compute, and, where possible, performance under a fixed total evaluation budget.
  3. [§3.2, Table 1 and Fig. 7] No error bars, seed-level variance, or significance tests are reported for the stochastic game tasks. Win-rates in Othello and Battleship are estimated from game samples, and the ablations in Fig. 7 are reported as single numbers. The claimed gaps (e.g., Othello-H 0.85 vs. ≤0.15 for baselines) are large, but without variance across seeds or evaluation replicates it is unclear whether the smaller differences in the easy variants, or the ablation deltas such as −0.06 on Othello-E, are meaningful. Please report mean ± std over multiple optimization seeds and, for evaluator noise, over repeated evaluations.
  4. [§2.3 and Appendix B, Eq. (8)] The manuscript alternates between exact subset enumeration (Eq. 7) and a budgeted subset-pool approximation (Eq. 8), but the experiments do not state, per domain, which mode was used, the number of generated edits n, the subset pool size, the budget B, or the number of evaluator calls assigned to each subset. This matters because the method's performance and cost depend critically on these choices: exact enumeration is infeasible for large n, and the budgeted approximation may behave differently. Please report these details for each task and include a sensitivity analysis of the budget B.
minor comments (5)
  1. [Appendix C.3] The text cross-references 'Table 2' for DeepSeek-v3.2 results, but that table appears in Appendix D (Extra Results). Please renumber or fix the cross-reference.
  2. [Eq. (7)] The normalization by ∥S∥ is set to 1 throughout, making it a no-op in the reported experiments. Please either remove it or explain what values ∥S∥ would take in practice and why the normalization is retained.
  3. [Fig. 7] The ablation table is difficult to parse as rendered: rows and columns should be labeled clearly, and the sign convention ('variant score minus full HERO score') should be stated directly in the figure caption.
  4. [§2.1, Eq. (2)] The latent-objective model g_1,...,g_h is presented as a 'tractable abstraction,' but X may be a discrete program space and ∇g_i is not defined. Please clarify that this is an informal intuition rather than a formal gradient, or provide a discrete analogue.
  5. [Appendix A] The terminology 'first-order information' is used in a nonstandard sense. The paper defines the term, but a brief remark contrasting it with numerical first-order optimization would help avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: HERO's empirical claims rest on external evaluations, and its theoretical statements are explicitly conditional, self-contained, and corrected where imprecise.

full rationale

The paper's central claim is empirical: HERO outperforms prior LLM optimizers on four external benchmarks. Nothing in the method fits a parameter to those benchmark outcomes and then reports the fit as a prediction. The selection rule (Eq. 7) uses the evaluator f only as a black-box scalar for ranking subsets; the LLM never sees f, so the zeroth-order label is substantive rather than definitional. The theoretical impossibility result in Appendix C is proved in-paper under explicit assumptions (A1)-(A4): with an unrestricted decoder and finite data, labels constrain only finitely many observed scores, so unidentifiability follows by construction. This is a conditional mathematical statement, not an imported self-citation or an empirically fitted premise. The latent-objective model (Eq. 2) and the additivity approximation (Eq. 4) are explicitly labeled as tractable abstractions rather than inferred facts, so even if Eq. (4) is empirically questionable, that is a correctness risk, not a circular derivation. The authors' flagged correction in Appendix C Remark 3 further shows the theoretical content is being stated carefully rather than assumed. Self-citations appear only in related-work examples (e.g., Fu et al. 2023, 2024; Li et al. 2025) and are not load-bearing for the main argument. No fitted input is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. The reported experiments are external and could have failed, so the derivation chain is self-contained.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

HERO does not fit parameters to data; the free parameters are hand-chosen cost/coverage budgets. The main axioms are the assumed latent-objective structure, the additivity of edit effects, the unrestricted decoder used in the impossibility theorem, and the broad premise that LLMs can propose useful edits without score feedback.

free parameters (2)
  • subset evaluation budget B
    Introduced in Eq. (8) to balance subset coverage vs. ranking reliability; chosen by hand per task, no concrete values are reported, so its effect on the empirical results is not pinned down.
  • surrogate evaluator granularity (number of replicate games)
    In Fig. 4A the sweep over replicate counts shows performance plateaus quickly, but the exact number used in final runs is not specified; this affects evaluation cost, not the fitted form of the method.
axioms (4)
  • ad hoc to paper Latent-objective model (Eq. 2): each sampled edit is an approximate ascent step on a random mixture of latent objectives g_1...g_h.
    Stated in Section 2.1 as 'a tractable abstraction, not a claim about internal mechanics'. It motivates Eq. (3) but is not empirically validated and is not required for the algorithm itself.
  • domain assumption Additive edit effects (Eq. 4) for non-overlapping edits.
    Load-bearing for the weakest-link analysis. Non-overlapping code regions are claimed to make effects approximately additive, but global objectives can couple disjoint edits (e.g., circle packing, game trajectories).
  • domain assumption Unrestricted decoder class (A1) in Appendix C.
    The impossibility theorem assumes all measurable decoders, which makes the unidentifiability result true by construction; the authors themselves note in Remark 1 that Lipschitz or parametric decoders restore identifiability.
  • domain assumption LLM prior contains plausible atomic edits without performance feedback.
    The central empirical premise of the Generation phase. Supported indirectly by the ablation results and cross-model experiments, but not guaranteed; the method would fail if the base LLM cannot propose sensible edits.
invented entities (1)
  • Latent objective functions g_1...g_h no independent evidence
    purpose: Formalize the unobserved 'qualities' the LLM's edits are assumed to ascend, enabling Eq. (2) and Eq. (3).
    Purely an analytical device introduced in Section 2.1; no independent measurement or falsifiable prediction is associated with these functions, and the empirical results do not depend on their existence.

pith-pipeline@v1.3.0-daily-deepseek · 248 in / 7297 out tokens · 107469 ms · 2026-08-03T16:40:57.752663+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) are increasingly used to solve complex problems by searching over program space, offering a general paradigm for scientific problems that can be naturally represented and solved as programs. Despite recent progress, identifying effective optimization directions for a candidate program remains challenging. By analogy with automatic differentiation, existing methods typically guide the search using a textual ``gradient'': a first-order update direction expressed as textual edits. Such gradients are inferred either from previously evaluated programs or from LLM-generated feedback on the implicit program-score mapping. However, these estimates become increasingly unreliable as the program--score mapping grows more complex, limiting their practical utility. We argue that explicit gradients are not essential for effective program optimization. Leveraging their prior knowledge, LLMs can propose plausible atomic edits directly from the current program, thereby enabling a zeroth-order optimization strategy. However, zeroth-order search suffers from a \textit{weakest-link effect}: when a bundle of edits is accepted or rejected as a whole, a single harmful edit can negate the benefits of all remaining edits. To address this issue, we introduce HERO, a program optimizer that prompts an LLM to generate diverse, non-overlapping atomic edits and then systematically selects and composes them into coherent program improvements using evaluator scores. We evaluate HERO across algorithmic problems, strategy games, the design of LLM-based agentic systems, and robotic path planning. Across these domains, HERO consistently discovers higher-scoring programs and converges substantially faster than prior LLM-based optimizers, while consuming fewer tokens.

Figures

Figures reproduced from arXiv: 2607.28947 by He Zhang, Jingwen Fu, Nanning Zheng, Yuhan Liu, Zhen Liu.

Figure 1
Figure 1. Figure 1: Workflow of HERO. Given a task description, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HERO. (A) Working mechanism. At each step, the LLM generates multiple non-overlapping, heteroge￾neous atomic edits. Exsiting methods apply or reject all edits by the global score, hence suffering from the weakest-link effect. HERO instead recombines the edits and selects the subset that maximizes performance, neutralizing the influence of harmful edits. (B) Best-subset recombination mitigates t… view at source ↗
Figure 3
Figure 3. Figure 3: Results on the Circle Packing task. A. Number of steps required to reach a configuration with sum_radii > 2.60. B. Average number of tokens consumed per evolution step. C. Optimization trajectories of different methods during the search. whole batch whenever one edit misfires, so progress com￾pounds across steps rather than stalling. This translates into strong sample efficiency, which matters most when in… view at source ↗
Figure 4
Figure 4. Figure 4: Analysis on Strategy Games. A. Effect of the surrogate evaluator E′ with different numbers of replicate games. B. Average performance of HERO, ShinkaEvolve, and OpenEvolve across three LLMs. Choice of Evaluator and Overfitting. A possible ob￾jection is that HERO simply trades expensive reasoning for expensive evaluation, since selection (Eq. 7) calls f repeat￾edly. In Othello, the faithful evaluator runs 1… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on Strategy Games. Each cell reports [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples of Battleship and Othello Games. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] 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

82 extracted references · 22 linked inside Pith

  1. [1]

    Alibaba. 2025. Qwen-Plus. https://qwen-ai.chat/models/qwen-plus/

    2025

  2. [2]

    Bosio, C.; and Mueller, M. W. 2025. Synthesizing interpretable control policies through large language model guided search. In 2025 American Control Conference (ACC), 583--590. IEEE

    2025

  3. [5]

    Cheng, C.-A.; Nie, A.; and Swaminathan, A. 2024. Trace is the next autodiff: Generative optimization with rich feedback, execution traces, and llms. Advances in Neural Information Processing Systems, 37: 71596--71642

    2024

  4. [7]

    Dor, S. 2018. Strategy in games or strategy games: Dictionary and encyclopaedic definitions for game studies. Game Studies, 18(1): 43--55

    2018

  5. [8]

    Fu, J.; Wang, B.; Zhang, H.; Zhang, Z.; Chen, W.; and Zheng, N. 2023. When and Why Momentum Accelerates SGD : An Empirical Study. arXiv:2306.09000

    Pith/arXiv arXiv 2023

  6. [9]

    Fu, J.; Zhang, X.; Wang, Y.; Zeng, W.; and Zheng, N. 2024. Understanding Mobile GUI : From Pixel-Words to Screen-Sentences. Neurocomputing, 601: 128200

    2024

  7. [10]

    Fu, J.; Zhu, X.; and Li, Y. 2019. Recognition of Surface Defects on Steel Sheet Using Transfer Learning. arXiv:1909.03258

    Pith/arXiv arXiv 2019

  8. [11]

    Guo, Q.; Wang, R.; Guo, J.; Li, B.; Song, K.; Tan, X.; Liu, G.; Bian, J.; and Yang, Y. 2024. Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers. In The Twelfth International Conference on Learning Representations

    2024

  9. [15]

    Li, P.; Wu, K.; Fu, J.; and Zhou, S. 2025. REGNav : Room Expert Guided Image-Goal Navigation. Proceedings of the AAAI Conference on Artificial Intelligence, 39(5): 4860--4868

    2025

  10. [17]

    Liu, Z.; Liu, Y.; Wang, J.; Liu, J.; Song, W.; and Fu, J. 2026 a . The Essence of Balance for Self-Improving Agents in Vision-and-Language Navigation. arXiv:2604.19064

    Pith/arXiv arXiv 2026

  11. [18]

    Liu, Z.; Liu, Y.; Wang, J.; Liu, J.; Song, W.; and Fu, J. 2026 b . Instruction-as-State: Environment-Guided and State-Conditioned Semantic Understanding for Embodied Navigation. arXiv:2604.18223

    Pith/arXiv arXiv 2026

  12. [20]

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36: 46534--46594

    2023

  13. [21]

    Mathematical Association of America . 2023. 2023 American Invitational Mathematics Examination ( AIME ) Problems and Solutions. American Mathematics Competitions

    2023

  14. [22]

    Mathematical Association of America . 2024. 2024 American Invitational Mathematics Examination ( AIME ) Problems and Solutions. American Mathematics Competitions

    2024

  15. [23]

    Mathematical Association of America . 2025. 2025 American Invitational Mathematics Examination ( AIME ) Problems and Solutions. American Mathematics Competitions

    2025

  16. [24]

    Meng, S.; Wang, Y.; Yang, C.; Peng, N.; and Chang, K. 2024. LLM-A*: Large Language Model Enhanced Incremental Heuristic Search on Path Planning

    2024

  17. [25]

    Nie, A.; Cheng, C.-A.; Kolobov, A.; and Swaminathan, A. 2023. The Importance of Directional Feedback for LLM -Based Optimizers. In NeurIPS 2023 Workshop on Foundation Models for Decision Making

    2023

  18. [27]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  19. [29]

    Romera-Paredes, B.; Little, D.; Torr, P.; et al. 2024. Mathematical discoveries from program search with large language models. Nature, 625(7999): 476--483

    2024

  20. [30]

    Shen, Y.; Zhou, S.; Fu, J.; Wang, R.; Chen, S.; and Zheng, N. 2023. StructVPR : Distill Structural Knowledge with Weighting Samples for Visual Place Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11217--11226

    2023

  21. [31]

    Wang, B. 2021. Path planning of mobile robot based on A* algorithm. In 2021 IEEE International Conference on Electronic Technology, Communication and Information (ICETCI), 524--528. IEEE

    2021

  22. [32]

    Wang, B.; Fu, J.; Zhang, H.; Zheng, N.; and Chen, W. 2023. Closing the Gap Between the Upper Bound and Lower Bound of Adam 's Iteration Complexity. In Advances in Neural Information Processing Systems, volume 36, 39006--39032

    2023

  23. [33]

    Wang, H.; Hao, C.; Zhang, P.; Zhang, M.; Yin, P.; and Zhang, Y. 2019. Path planning of mobile robots based on A* algorithm and artificial potential field algorithm. China mechanical engineering, 30(20): 2489

    2019

  24. [36]

    V.; Zhou, D.; and Chen, X

    Yang, C.; Wang, X.; Lu, Y.; Liu, H.; Le, Q. V.; Zhou, D.; and Chen, X. 2023. Large language models as optimizers. In The Twelfth International Conference on Learning Representations

    2023

  25. [37]

    E.; Wettig, A.; Lieret, K.; Yao, S.; Narasimhan, K.; and Press, O

    Yang, J.; Jimenez, C. E.; Wettig, A.; Lieret, K.; Yao, S.; Narasimhan, K.; and Press, O. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems, 37: 50528--50652

    2024

  26. [38]

    Ye, H.; Wang, J.; Cao, Z.; Berto, F.; Hua, C.; Kim, H.; Park, J.; and Song, G. 2024. Reevo: Large language models as hyper-heuristics with reflective evolution. Advances in neural information processing systems, 37: 43571--43608

    2024

  27. [39]

    Yuksekgonul, M.; Bianchi, F.; Boen, J.; Liu, S.; Lu, P.; Huang, Z.; Guestrin, C.; and Zou, J. 2025. Optimizing generative ai by backpropagating language model feedback. Nature, 639(8055): 609--616

    2025

  28. [40]

    Zhang, J.; Xiang, J.; Yu, Z.; Teng, F.; Chen, X.-H.; Chen, J.; Zhuge, M.; Cheng, X.; Hong, S.; Wang, J.; et al. 2025 a . AFlow : Automating Agentic Workflow Generation. In The Thirteenth International Conference on Learning Representations

    2025

  29. [41]

    Zhang, P.; Jin, H.; Hu, L.; Li, X.; Kang, L.; Luo, M.; Song, Y.; and Wang, H. 2024 a . HessianGrad : Optimizing AI Systems with Hessian-Aware Textual Gradients. ICLR 2025 conference submission

    2024

  30. [43]

    Zhang, Y.; Ruan, H.; Fan, Z.; and Roychoudhury, A. 2024 c . Autocoderover: Autonomous program improvement. In Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 1592--1604

    2024

  31. [46]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

  32. [47]

    Nature , volume=

    Optimizing generative ai by backpropagating language model feedback , author=. Nature , volume=. 2025 , publisher=

    2025

  33. [48]

    The Twelfth International Conference on Learning Representations , year=

    Large language models as optimizers , author=. The Twelfth International Conference on Learning Representations , year=

  34. [49]

    The Thirteenth International Conference on Learning Representations , year=

    Efficient Evolutionary Search over Chemical Space with Large Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  35. [50]

    Nature , volume=

    Mathematical discoveries from program search with large language models , author=. Nature , volume=. 2024 , publisher=

    2024

  36. [51]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  37. [52]

    arXiv preprint arXiv:2406.04824 , year=

    Funbo: Discovering acquisition functions for bayesian optimization with funsearch , author=. arXiv preprint arXiv:2406.04824 , year=

    Pith/arXiv arXiv

  38. [53]

    2025 , url=

    Zhang, Jiayi and Xiang, Jinyu and Yu, Zhaoyang and Teng, Fengwei and Chen, Xiong-Hui and Chen, Jiaqi and Zhuge, Mingchen and Cheng, Xin and Hong, Sirui and Wang, Jinlin and others , booktitle=. 2025 , url=

    2025

  39. [54]

    arXiv preprint arXiv:2408.08435 , year=

    Automated design of agentic systems , author=. arXiv preprint arXiv:2408.08435 , year=

    Pith/arXiv arXiv

  40. [55]

    arXiv preprint arXiv:2506.13131 , year=

    AlphaEvolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

    Pith/arXiv arXiv

  41. [56]

    2025 , howpublished=

    2025 American Invitational Mathematics Examination (. 2025 , howpublished=

    2025

  42. [57]

    2024 , howpublished=

    2024 American Invitational Mathematics Examination (. 2024 , howpublished=

    2024

  43. [58]

    2023 , howpublished=

    2023 American Invitational Mathematics Examination (. 2023 , howpublished=

    2023

  44. [59]

    AMC contests , year =

    the Mathematical Association of America , title =. AMC contests , year =

  45. [60]

    arXiv preprint arXiv:2509.19349 , year=

    Shinkaevolve: Towards open-ended and sample-efficient program evolution , author=. arXiv preprint arXiv:2509.19349 , year=

    Pith/arXiv arXiv

  46. [61]

    arXiv preprint arXiv:2412.03092 , year=

    Revolve: Optimizing ai systems by tracking response evolution in textual optimization , author=. arXiv preprint arXiv:2412.03092 , year=

    Pith/arXiv arXiv

  47. [62]

    The Importance of Directional Feedback for

    Nie, Allen and Cheng, Ching-An and Kolobov, Andrey and Swaminathan, Adith , booktitle=. The Importance of Directional Feedback for. 2023 , url=

    2023

  48. [63]

    Advances in Neural Information Processing Systems , volume=

    Trace is the next autodiff: Generative optimization with rich feedback, execution traces, and llms , author=. Advances in Neural Information Processing Systems , volume=

  49. [64]

    arXiv preprint arXiv:2506.00400 , year=

    Scaling Textual Gradients via Sampling-Based Momentum , author=. arXiv preprint arXiv:2506.00400 , year=

    Pith/arXiv arXiv

  50. [65]

    Nature , year =

    Romera-Paredes, Bernardino and Little, Daniel and Torr, Philip and others , title =. Nature , year =. doi:10.1038/s41586-023-06924-6 , url =

    doi:10.1038/s41586-023-06924-6

  51. [66]

    Game Studies , volume=

    Strategy in games or strategy games: Dictionary and encyclopaedic definitions for game studies , author=. Game Studies , volume=

  52. [67]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Yao, Shuhan and Zhou, Yuxuan and Zhang, Yuxuan and Yu, Xiangliang , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =. doi:10.1609/aaai.v39i25.34922 , url =

    doi:10.1609/aaai.v39i25.34922

  53. [68]

    arXiv preprint arXiv:2401.02051 , year=

    Evolution of heuristics: Towards efficient automatic algorithm design using large language model , author=. arXiv preprint arXiv:2401.02051 , year=

    Pith/arXiv arXiv

  54. [69]

    arXiv preprint arXiv:2403.02054 , year =

    Large Language Model-Based Evolutionary Optimizer: Reasoning with Elitism , author =. arXiv preprint arXiv:2403.02054 , year =

    Pith/arXiv arXiv

  55. [70]

    arXiv preprint arXiv:2507.14172 , year=

    Self-improving language models for evolutionary program synthesis: A case study on ARC-AGI , author=. arXiv preprint arXiv:2507.14172 , year=

    arXiv

  56. [71]

    arXiv preprint arXiv:2312.06853 , year=

    Llf-bench: Benchmark for interactive learning from language feedback , author=. arXiv preprint arXiv:2312.06853 , year=

    Pith/arXiv arXiv

  57. [72]

    arXiv preprint arXiv:2508.19506 , year=

    Learning Game-Playing Agents with Generative Code Optimization , author=. arXiv preprint arXiv:2508.19506 , year=

    Pith/arXiv arXiv

  58. [73]

    2025 American Control Conference (ACC) , pages=

    Synthesizing interpretable control policies through large language model guided search , author=. 2025 American Control Conference (ACC) , pages=. 2025 , organization=

    2025

  59. [74]

    arXiv preprint arXiv:2508.10059 , year=

    CodeGrad: Integrating Multi-Step Verification with Gradient-Based LLM Refinement , author=. arXiv preprint arXiv:2508.10059 , year=

    Pith/arXiv arXiv

  60. [75]

    Advances in Neural Information Processing Systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in Neural Information Processing Systems , volume=

  61. [76]

    arXiv preprint arXiv:2506.16389 , year=

    RiOT: Efficient Prompt Refinement with Residual Optimization Tree , author=. arXiv preprint arXiv:2506.16389 , year=

    Pith/arXiv arXiv

  62. [77]

    Advances in Neural Information Processing Systems , volume=

    Swe-agent: Agent-computer interfaces enable automated software engineering , author=. Advances in Neural Information Processing Systems , volume=

  63. [78]

    Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis , pages=

    Autocoderover: Autonomous program improvement , author=. Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis , pages=

  64. [79]

    arXiv preprint arXiv:2407.01489 , year=

    Agentless: Demystifying llm-based software engineering agents , author=. arXiv preprint arXiv:2407.01489 , year=

    Pith/arXiv arXiv

  65. [80]

    The Thirteenth International Conference on Learning Representations , year=

    Automated Design of Agentic Systems , author=. The Thirteenth International Conference on Learning Representations , year=

  66. [81]

    LLM-A*: Large Language Model Enhanced Incremental Heuristic Search on Path Planning , author=

  67. [82]

    Zhang, Peiyan and Jin, Haibo and Hu, Leyang and Li, Xinnuo and Kang, Liying and Luo, Man and Song, Yangqiu and Wang, Haohan , year=

  68. [83]

    arXiv preprint arXiv:2312.01797 , year=

    Llm a*: Human in the loop large language models enabled a* search for robotics , author=. arXiv preprint arXiv:2312.01797 , year=

    Pith/arXiv arXiv

  69. [84]

    The Twelfth International Conference on Learning Representations , year=

    Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers , author=. The Twelfth International Conference on Learning Representations , year=

  70. [85]

    Advances in neural information processing systems , volume=

    Reevo: Large language models as hyper-heuristics with reflective evolution , author=. Advances in neural information processing systems , volume=

  71. [86]

    arXiv preprint arXiv:1805.08137 , year=

    Robot path planning algorithms: a review of theory and experiment , author=. arXiv preprint arXiv:1805.08137 , year=

    Pith/arXiv arXiv

  72. [87]

    China mechanical engineering , volume=

    Path planning of mobile robots based on A* algorithm and artificial potential field algorithm , author=. China mechanical engineering , volume=

  73. [88]

    2021 IEEE International Conference on Electronic Technology, Communication and Information (ICETCI) , pages=

    Path planning of mobile robot based on A* algorithm , author=. 2021 IEEE International Conference on Electronic Technology, Communication and Information (ICETCI) , pages=. 2021 , organization=

    2021

  74. [89]

    arXiv preprint arXiv:2605.09271 , year=

    Shaping schema via language representation as the next frontier for llm intelligence expanding , author=. arXiv preprint arXiv:2605.09271 , year=

    Pith/arXiv arXiv

  75. [90]

    arXiv preprint arXiv:2605.04057 , year=

    Structured Progressive Knowledge Activation for LLM-Driven Neural Architecture Search , author=. arXiv preprint arXiv:2605.04057 , year=

    Pith/arXiv arXiv

  76. [91]

    2026 , eprint =

    Liu, Zhen and Liu, Yuhan and Wang, Jinjun and Liu, Jianyi and Song, Wei and Fu, Jingwen , title =. 2026 , eprint =

    2026

  77. [92]

    Advances in Neural Information Processing Systems , volume =

    Wang, Bohan and Fu, Jingwen and Zhang, Huishuai and Zheng, Nanning and Chen, Wei , title =. Advances in Neural Information Processing Systems , volume =. 2023 , doi =

    2023

  78. [93]

    2019 , eprint =

    Fu, Jingwen and Zhu, Xiaoyan and Li, Yingbin , title =. 2019 , eprint =

    2019

  79. [94]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Shen, Yanqing and Zhou, Sanping and Fu, Jingwen and Wang, Ruotong and Chen, Shitao and Zheng, Nanning , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2023 , doi =

    2023

  80. [95]

    2023 , eprint =

    Fu, Jingwen and Wang, Bohan and Zhang, Huishuai and Zhang, Zhizheng and Chen, Wei and Zheng, Nanning , title =. 2023 , eprint =

    2023

Showing first 80 references.