Pith. sign in

REVIEW 3 major objections 5 minor 30 references

This paper claims that agentic workflow synthesis can be done at test time, without any training, by jointly searching over a task-decomposition topology and executable sub-workflows, with execution feedback able to reshape the topology its

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

2026-08-02 13:32 UTC pith:IPPLS4FT

load-bearing objection HierFlow's coupled topology–execution search and gating knob are genuinely new, and the paper is unusually honest about its own limits, but the empirical 'consistently outperforms' claim needs paired statistics and cleaner baselines before it lands. the 3 major comments →

arxiv 2607.21609 v1 pith:IPPLS4FT submitted 2026-05-20 cs.AI

Coupled Hierarchical Search over Topology and Execution for Agentic Workflow Synthesis

Dong Li , Yanchi Liu , Xujiang Zhao , Wei Cheng , Zhengzhang Chen , Xintao Wu , Zhong Chen , Chen Zhao
show 1 more author
Haifeng Chen
This is my paper
classification cs.AI
keywords agentic workflow synthesistest-time scalinghierarchical searchMCTS-inspired searchgating mechanismtopology refinementLLM-based evaluationtraining-free optimization
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 is trying to establish that automating agentic workflow design can be solved at test time and without any training by searching over two coupled spaces: a task-decomposition topology and executable sub-workflows. It claims that execution feedback should not only refine the code within a subtask but also revise the decomposition itself, making the topology a search variable rather than a fixed pipeline. The proposed system, HierFlow, is claimed to outperform both handcrafted reasoning methods and automatic workflow-generation baselines—including ones that required offline training—on question answering, math reasoning, and code generation benchmarks, while offering a threshold that trades computational cost for quality. A sympathetic reader would care because this points to a way to get adaptive, query-level workflows without the data-collection and training overhead of existing approaches.

Core claim

On the paper's own terms, the central discovery is that the workflow itself is a coupled structure–execution state: a directed acyclic graph of subtasks together with an executable sub-workflow attached to each node. Both components are searched simultaneously: the topology defines the code-search domain for each subtask, and failures or low confidence in the lower-level execution can trigger a topology revision that adds, removes, or rewires subtasks and rewinds execution. HierFlow instantiates this with an LLM-driven upper-level refiner, an MCTS-inspired lower-level tree search with execution-grounded evaluation, and a gating mechanism that activates lower-level search only when a search-n

What carries the argument

The load-bearing machinery is the coupled dual-space proxy: an upper-level topology space of DAG task decompositions and a lower-level space of executable code sub-workflows, searched jointly. Three mechanisms carry the argument: (i) a LLM-based topology refiner that scores candidate decompositions with a weighted objective including a dependency-complexity term, (ii) a lightweight MCTS-inspired tree search that iterates selection, expansion, evaluation (including a hard sandbox executability check), and information propagation without score backpropagation to parents, and (iii) a gating module that uses the search necessity proxy Z_i = S_i * U_i, where S_i is log remaining candidate space a

Load-bearing premise

The central load-bearing premise is that the LLM-based evaluator's quality scores are a faithful proxy for end-task correctness when ranking candidate sub-workflows and when computing the gating signal; if this alignment weakens on new tasks or distributions, both the lower-level search and the gating decision will select suboptimal workflows.

What would settle it

A concrete experiment that would settle the claim: on a new distribution, compute the Spearman correlation between the gating proxy Z_i and the realized quality gain from forcing lower-level search; if this correlation is not significantly positive, or if the stratified improvement rate does not increase with Z_i, the gating mechanism's justification fails. Likewise, on a task with deliberately high cross-subtask coupling (e.g., steps that must share and update global state), the predicted degradation should appear: the system's cost–quality trade-off should cease to beat a flat MCTS-in-code b

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

If this is right

  • If the central claim holds, automatic workflow generation no longer requires offline search over a fixed dataset or training a designer; a query-level, training-free search can match or exceed those methods.
  • The gating threshold τ gives users an interpretable control knob for test-time scaling: smaller τ buys accuracy with more compute, larger τ saves cost, and the operating point near the knee of the Pareto curve is competitive with the best baseline.
  • Execution feedback into topology revision means the system can recover from poor initial decompositions mid-query, which is a capability that fixed-topology or offline-searched workflows lack.
  • The coupling-aware analysis predicts that the advantage of hierarchical splitting shrinks as cross-subtask dependency density or repair rate grows, so the method's target regime is moderately decomposable tasks.
  • LLM-based evaluation, despite its known pitfalls, is claimed to be reliable enough for search guidance: on MATH the lower-level evaluator agrees with end-task success in 85.8% of cases, and swapping judge models shifts accuracy by about one point.

Where Pith is reading between the lines

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

  • The rank-based gating theorem suggests a general recipe: any proxy that ranks subtasks by expected search gain can be plugged into the same budget-allocation framework. Z_i = S_i * U_i is one instantiation; a natural extension would be to use a cheap learned surrogate of realized gain, or an ensemble of uncertainty measures, without giving up the training-free property.
  • The observable coupling proxy could be used as a per-query diagnostic: compute it before or during search to decide when to switch to flat code search or to adjust the gating threshold, potentially making the method robust to tasks that violate the moderate-coupling assumption.
  • Because the lower-level search is execution-grounded with mock inputs, the method transfers to any domain with a sandbox and a schema-consistent input generator—for example formal proof verification, SQL query generation, or tool-use workflows where execution feedback is available. This is a testable extension, not something the paper claims.
  • The memory bank that reuses sub-workflow configurations across semantically similar subtasks suggests a form of amortized test-time search; across a stream of queries, the system could accumulate reusable components without training, which the paper does not explore.

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 / 5 minor

Summary. HierFlow is proposed as a training-free, test-time hierarchical search framework for agentic workflow synthesis. It searches jointly over an upper-level task-topology space and a lower-level code-execution space, with a gating mechanism that decides when to invoke execution-level search. The paper claims consistent state-of-the-art results on HotpotQA, GSM8K, MATH, MBPP, and HumanEval, and supports the design with a coupling-aware cost analysis, a rank-based gating analysis, ablations, and evaluator-robustness checks.

Significance. If the central empirical claim holds, HierFlow is a useful contribution: it combines two complementary proxy spaces in a training-free manner, offers a tunable quality–cost trade-off through the gating threshold, and includes several honest robustness analyses (judge-model sensitivity, evaluator–end-task agreement, direct gating-proxy validation). The theoretical results are correctly framed as cost-accounting and rank-bound statements rather than unconditional guarantees. However, the main empirical claim of consistent superiority is currently supported only by point estimates with overlapping standard deviations, and one baseline is confounded by a different base model. These issues are fixable but load-bearing.

major comments (3)
  1. [§3.2, Table 1] The central claim of consistent superiority rests on point estimates whose headline margins are within one standard deviation of the runner-up on three of five benchmarks: GSM8K 92.70±0.35 vs MaAS 92.51±0.39 (Δ=0.19), MBPP 82.40±0.77 vs MaAS 82.11±0.84 (Δ=0.29), and HumanEval 93.13±0.88 vs DyFlow 92.37±0.79 (Δ=0.76). With only three aggregate runs and no paired significance tests, confidence intervals, or effect sizes, the abstract's 'consistently outperforms' is under-evidenced. Please add per-example paired tests (e.g., bootstrap or Wilcoxon signed-rank over the test sets of size n=800/1055/486/131/341) and report CIs; otherwise the claim should be weakened.
  2. [§3.1/Appendix D.2, Table 1] The comparison to DyFlow is confounded by base model: 'The only exception is DyFlow, for which we directly adopt its pre-trained designer based on Phi-4' (Appendix D.2), while HierFlow and all other baselines use gpt-4o-mini. DyFlow is the runner-up on MATH and HumanEval, the two benchmarks where the claimed advantage is largest. Since base-model capability can explain part of this gap, the method-level comparison requires either a DyFlow variant with a gpt-4o-mini designer or a Phi-4-based HierFlow run. Without this, the headline margins over DyFlow are not attributable to the search mechanism.
  3. [Appendix D.5, Definition 2.5] The validation of the gating proxy is partially self-referential: ΔQ_i is computed as Q_i(after lower-level search) − Q_i(baseline IO), where Q_i is the same LLM evaluator Jsub used by the lower-level search, and Z_i uses the predictive entropy of rationales produced by the same evaluator. A positive Spearman ρ=0.43 may therefore reflect evaluator self-consistency rather than alignment with end-task correctness. Please report the rank correlation of Z_i with externally measured gains (e.g., final pass@1/accuracy on the test examples) or with a different judge model, and stratify improvement rates by external correctness. The current 85.8% evaluator–end-task agreement (Table 7) does not directly establish proxy informativeness.
minor comments (5)
  1. [§3.4, Table 2] The header 'Training ↓ Testing MATH HumanEval GSM8K MBPP HotpotQA' is ambiguous, and the HotpotQA values for DyFlow (68.13) and Flow (67.88) differ from Table 1 (67.88 and 67.63). Please clarify the transfer protocol and reconcile the numbers.
  2. [§2.3] Typo: 'he search process' should read 'The search process'.
  3. [References] References [24] and [25] appear to be the same citation; [25] should cite the Self-Consistency paper, not the CoT paper.
  4. [§3.1] The text says 'Results average three independent runs' but does not report seeds or explain the variance source. Please add this information or per-run results for the main table.
  5. [Figure 3] Minor formatting: 'MATHdataset' should be 'MATH dataset'; the 'A vg.' label in Table 1 should be 'Avg.'

Circularity Check

1 steps flagged

Gating-proxy validation is partly self-referential, but the headline benchmark results are external and not circular.

specific steps
  1. other [Appendix D.5 (Empirical Validation of the Gating Proxy), combined with Sec. 2.3 Evaluation and Definition 2.4]
    "For each candidate node encountered during upper-level traversal, we record its proxy value Zi before the gating decision and then force the lower-level search to be invoked under a fixed budget, regardless of whether the gating rule would have triggered it. We measure the realized quality gain ΔQi = Qi(after lower-level search)−Qi(baseline IO solution), where the baseline corresponds to the root node produced by direct input–output generation."

    The gating proxy Zi = Si·Ui includes Ui, the predictive entropy of the evaluation rationale ri produced by the LLM-based evaluator Jsub, and the same Jsub assigns the quality scores whose difference defines ΔQi. The reported Spearman ρ=0.43 and the stratified improvement rates therefore partly measure whether the evaluator's own rationale uncertainty predicts its own score gain, rather than independently validating that gating selects subtasks that improve end-task correctness. This is not full circularity because Zi and ΔQi are distinct quantities and Table 1 reports external pass@1/F1/accuracy, but it is a self-referential supporting validation.

full rationale

I find no step in which the paper's central benchmark claim reduces by construction to its inputs. Theorem 2.2 is an explicit bookkeeping decomposition of test-time cost, and the paper itself labels it a 'coupling-aware characterization rather than an unconditional asymptotic speedup guarantee' (Remark 2.3). Theorem 2.6 is a rank-disagreement bound that explicitly disclaims unconditional near-optimality. The final Table 1 numbers are scored by external metrics (F1, accuracy, pass@1), so the main 'consistently outperforms' claim is not circular in the sense of being equivalent to a fitted quantity. The one genuine circularity concern is the gating validation in Appendix D.5: the realized gain ΔQi is computed from the same LLM evaluator that supplies the uncertainty component of the gating proxy Zi, so the reported rank correlation and improvement rates are partly self-referential. The paper partially mitigates this with judge-model sensitivity and end-task-agreement analyses (Appendix D.4), and the gating validation is supporting rather than load-bearing for the headline result. There is also a minor self-citation in the introduction ('test-time scaling has emerged ... [16]', where [16] is the authors' own SolverLLM), but this is not load-bearing because the test-time scaling paradigm is independently established and the method's design does not rest on that specific prior result. Overall, the derivation is largely self-contained, with one partial self-referential validation, corresponding to a score of 4.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The paper's central contribution is an empirical system plus an honest analysis for when the hierarchy pays off. It introduces several hand-set hyperparameters and relies on LLM-evaluator fidelity as a domain assumption; no new external entities are postulated beyond a gating heuristic that is validated mostly internally.

free parameters (7)
  • Gating threshold τ = 4
    Selected as the operating point on the MATH cost–accuracy Pareto front (Fig. 3); controls how often lower-level search is triggered.
  • Refinement signal threshold γ (Eq. 3) = not specified
    Dichotomizes subtasks into refinement-needed vs not; sensitivity is not reported.
  • Lower-level selection parameters α, λp = α=0.4, λp=0.3
    Controls exploration/exploitation in the MCTS-inspired selection; chosen by hand.
  • Topology scoring weights λ0, λcov, λsnd, λsea = uniform
    Set uniformly; no sensitivity analysis.
  • Coupling proxy weights β, η (Eq. 1) = unspecified
    Weights of degree imbalance and repair-rate terms in κ̂; never instantiated in experiments.
  • Memory bank thresholds α=0.5, θ=0.85 = 0.5, 0.85
    Similarity weighting and retrieval threshold for sub-workflow reuse; chosen without sensitivity analysis.
  • Search budgets dmax=3, bmax=3, ccmax=5, K=3, ctmax=20 = 3,3,5,3,20
    Budget hyperparameters; K and ccmax directly set the cost and depth of search.
axioms (5)
  • domain assumption LLM evaluator quality scores align with end-task correctness
    The lower-level search selects sub-workflows by an LLM judge (Sec. 2.3); the paper verifies 85.8% agreement on MATH (Appendix D.4.3) but assumes this holds on all benchmarks and distributions.
  • domain assumption Topological layer-wise execution is a valid decomposition for the target tasks
    HierFlow assumes tasks can be decomposed into DAG-structured subtasks with localized code search spaces; this is a modeling assumption borrowed from Flow [20].
  • domain assumption The search-space estimate |W_i| ≈ b_max^{d_max} governs search potential
    The gating formula uses this estimate to define search potential S_i; it is a rough proxy, not a proven identity.
  • standard math Standard combinatorial and statistical results (Kahn's algorithm, Softmax selection, Spearman correlation)
    Used as black-box tools; no new mathematics is introduced.
  • ad hoc to paper Observable coupling proxy κ̂(G) is a meaningful measure of cross-task coupling
    Definition 2.1 is introduced ad hoc; it mixes structural and execution-observed terms with unspecified weights β, η, and is not derived from a task model.
invented entities (1)
  • Search necessity proxy Z_i = S_i · U_i no independent evidence
    purpose: Decides whether to trigger lower-level execution search for subtask i (gating).
    It is a heuristic construct; its validation in Appendix D.5 uses the paper's own LLM evaluator for both the proxy and the measured gain, so the evidence is not independent of the construct.

pith-pipeline@v1.3.0-alltime-deepseek · 20123 in / 16114 out tokens · 165492 ms · 2026-08-02T13:32:05.709423+00:00 · methodology

0 comments
read the original abstract

Although structured workflows empower Large Language Models (LLMs) to tackle complex problems, automating their creation is severely hindered by a vast combinatorial search space, frequently resulting in inflexible and resource-heavy offline training dependencies. To address this, we conceptualize workflow generation as an intertwined topology-and-execution search paradigm, where the broader topological layer dictates subtask boundaries and lower-level execution outcomes actively reshape the topology itself. Building on this foundation, we introduce HierFlow, a training-free, test-time hierarchical search architecture that automates agentic workflow design by merging feedback-guided topology adjustments with a fast, MCTS-inspired tree search for sub-workflow optimization. HierFlow maximizes efficiency through an intelligent gating module that selectively triggers execution-level searches based on contextual necessity, a mechanism we further support with an in-depth analysis detailing how varying degrees of cross-task coupling impact the effectiveness of hierarchical splitting. Comprehensive testing across question answering, mathematical reasoning, and code generation benchmarks confirms that HierFlow consistently outperforms strong baselines, delivering an optimal balance of high-quality results and computational efficiency without any additional training overhead.

Figures

Figures reproduced from arXiv: 2607.21609 by Chen Zhao, Dong Li, Haifeng Chen, Wei Cheng, Xintao Wu, Xujiang Zhao, Yanchi Liu, Zhengzhang Chen, Zhong Chen.

Figure 1
Figure 1. Figure 1: Three equivalent workflow representa￾tions (proxies) for a simple mathematical question. Although these methods surpass the per￾formance of manually designed systems, they mainly generate task-level workflows through offline search [11, 31] or query-level workflows via training a designer [23, 30], which not only incur substantial data and engineering costs but also remain tied to fixed training distri￾but… view at source ↗
Figure 2
Figure 2. Figure 2: Illustrative example of how HierFlow operates at inference time. Given a query, HierFlow [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pareto fronts illustrating the trade-off between [PITH_FULL_IMAGE:figures/full_fig_p010_3.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

30 extracted references · 14 linked inside Pith

  1. [1]

    Abdin, J

    M. Abdin, J. Aneja, H. Behl, S. Bubeck, R. Eldan, S. Gunasekar, M. Harrison, R. J. Hewett, M. Javaheripi, P. Kauffmann, et al. Phi-4 technical report.arXiv preprint arXiv:2412.08905, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    D. B. Acharya, K. Kuppan, and B. Divya. Agentic ai: Autonomous intelligence for complex goals–a comprehensive survey.IEEe Access, 2025

    2025

  3. [3]

    Astorga, T

    N. Astorga, T. Liu, Y. Xiao, and M. van der Schaar. Autoformulation of mathematical optimization models using llms.arXiv preprint arXiv:2411.01679, 2024. 11

    Pith/arXiv arXiv 2024

  4. [4]

    Austin, A

    J. Austin, A. Odena, M. Nye, M. Bosma, H. Michalewski, D. Dohan, E. Jiang, C. Cai, M. Terry, Q. Le, et al. Program synthesis with large language models.arXiv preprint arXiv:2108.07732, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Besta, N

    M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682–17690, 2024

    2024

  6. [6]

    M. Chen, J. Tworek, H. Jun, et al. Evaluating large language models trained on code. 2021

    2021

  7. [7]

    Cobbe, V

    K. Cobbe, V. Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch. Improving factuality and reasoning in language models through multiagent debate. InForty-first International Conference on Machine Learning, 2023

    2023

  9. [9]

    J. Fang, Y. Peng, X. Zhang, Y. Wang, X. Yi, G. Zhang, Y. Xu, B. Wu, S. Liu, Z. Li, et al. A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems.arXiv preprint arXiv:2508.07407, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Hendrycks, C

    D. Hendrycks, C. Burns, S. Kadavath, A. Arora, S. Basart, E. Tang, D. Song, and J. Stein- hardt. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874, 2021

    Pith/arXiv arXiv 2021

  11. [11]

    S. Hu, C. Lu, and J. Clune. Automated design of agentic systems.arXiv preprint arXiv:2408.08435, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Hughes, Y

    L. Hughes, Y. K. Dwivedi, T. Malik, M. Shawosh, M. A. Albashrawi, I. Jeon, V. Dutot, M. Appanderanda, T. Crick, R. De’, et al. Ai agents and agentic systems: A multi-expert analysis.Journal of Computer Information Systems, pages 1–29, 2025

    2025

  13. [13]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Kadavath, T

    S. Kadavath, T. Conerly, A. Askell, T. Henighan, D. Drain, E. Perez, N. Schiefer, Z. Hatfield- Dodds, N. DasSarma, E. Tran-Johnson, et al. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    A. B. Kahn. Topological sorting of large networks.Communications of the ACM, 5(11):558–562, 1962

    1962

  16. [16]

    D. Li, X. Zhao, L. Yu, Y. Liu, W. Cheng, Z. Chen, Z. Chen, F. Chen, C. Zhao, and H. Chen. Solverllm: Leveraging test-time scaling for optimization problem via llm-guided search.arXiv preprint arXiv:2510.16916, 2025

    arXiv 2025

  17. [17]

    X. Li, S. Wang, S. Zeng, Y. Wu, and Y. Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    2024

  18. [18]

    Z. Li, S. Xu, K. Mei, W. Hua, B. Rama, O. Raheja, H. Wang, H. Zhu, and Y. Zhang. Autoflow: Automated workflow generation for large language model agents.arXiv preprint arXiv:2407.12821, 2024. 12

    Pith/arXiv arXiv 2024

  19. [19]

    Madaan, N

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

    2023

  20. [20]

    B. Niu, Y. Song, K. Lian, Y. Shen, Y. Yao, K. Zhang, and T. Liu. Flow: Modularized agentic workflow automation.arXiv preprint arXiv:2501.07834, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    S. Qiao, R. Fang, Z. Qiu, X. Wang, N. Zhang, Y. Jiang, P. Xie, F. Huang, and H. Chen. Benchmarking agentic workflow generation.arXiv preprint arXiv:2410.07869, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Singh, A

    A. Singh, A. Ehtesham, S. Kumar, and T. T. Khoei. Enhancing ai systems with agentic workflows patterns in large language model. In2024 IEEE World AI IoT Congress (AIIoT), pages 527–532. IEEE, 2024

    2024

  23. [23]

    Y. Wang, Z. Xu, Y. Huang, X. Wang, Z. Song, L. Gao, C. Wang, X. Tang, Y. Zhao, A. Co- han, et al. Dyflow: Dynamic workflow framework for agentic reasoning.arXiv preprint arXiv:2509.26062, 2025

    arXiv 2025

  24. [25]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  25. [26]

    S. Xu, J. Zhang, S. Di, Y. Luo, L. Yao, H. Liu, J. Zhu, F. Liu, and M.-L. Zhang. Robustflow: Towards robust agentic workflow generation.arXiv preprint arXiv:2509.21834, 2025

    arXiv 2025

  26. [27]

    Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–2380, 2018

    2018

  27. [28]

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y. Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    2023

  28. [29]

    C. Yu, Z. Cheng, H. Cui, Y. Gao, Z. Luo, Y. Wang, H. Zheng, and Y. Zhao. A survey on agent workflow–status and future. In2025 8th International Conference on Artificial Intelligence and Big Data (ICAIBD), pages 770–781. IEEE, 2025

    2025

  29. [30]

    Zhang, L

    G. Zhang, L. Niu, J. Fang, K. Wang, L. Bai, and X. Wang. Multi-agent architecture search via agentic supernet.arXiv preprint arXiv:2502.04180, 2025

    Pith/arXiv arXiv 2025

  30. [31]

    Feedback- Driven

    J. Zhang, J. Xiang, Z. Yu, F. Teng, X. Chen, J. Chen, M. Zhuge, X. Cheng, S. Hong, J. Wang, et al. Aflow: Automating agentic workflow generation.arXiv preprint arXiv:2410.10762, 2024. 13 A Related Work Workflow Representation via Proxy Spaces.Since direct search over the workflow space is highly challenging, existing approaches typically reformulate workf...

    Pith/arXiv arXiv 2024