pith. sign in

arxiv: 2606.12674 · v1 · pith:ANDWZGQVnew · submitted 2026-06-10 · 💻 cs.AI

Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents

Kushal Raj Bhandari , Ling Yue , Ching-Yun Ko , Dhaval Patel , Shaowu Pan , Pin-Yu Chen , Jianxi Gao This is my paper

Pith reviewed 2026-06-27 09:45 UTC · model grok-4.3

classification 💻 cs.AI
keywords evolutionary searchtool workflowscompact agentsinference-time optimizationexecution feedbackMCP-Benchworkflow repairagent planning
0
0 comments X

The pith

Evoflux evolves executable tool workflows at inference time to raise small planners' execution feasibility from 3% to 17-24%.

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

Compact language models reduce cost for tool agents but frequently produce workflow graphs that fail tool resolution, schema checks, dependency tracking, or execution. Limited teacher traces teach format but leave recovery from failures uncovered when tool catalogs change. Evoflux performs evolutionary search over typed workflow graphs at inference time, applying structured edits, execution feedback, adaptive intensity, meta-guided redesign, and diversity pruning to repair plans. On held-out MCP-Bench tasks with live servers and 250 tools, this lifts feasibility to 17-24% across small planners. Supervised fine-tuning and DPO on the same mined data match or fall below zero-shot baselines, while ReAct reaches higher peaks at greater token cost and variance.

Core claim

Evoflux treats compact tool use as the repair of executable tool workflows. It evolves typed workflow graphs through structured edits, execution feedback, adaptive intensity, meta-guided redesign, and diversity pruning. On held-out MCP-Bench tasks spanning live MCP servers and 250 tools, Evoflux raises execution feasibility from roughly 3% to 17-24% across small planners. In contrast, SFT and SFT+DPO on the same search-mined data match, underperform, or collapse below zero-shot performance; ReAct reaches higher peaks, but with higher variance and token cost.

What carries the argument

Evolutionary search over typed workflow graphs using structured edits and execution feedback.

Load-bearing premise

Execution feedback from live tool calls, together with the structured edits and diversity pruning, suffices to discover repairs that generalize across changing tool catalogs without more teacher traces.

What would settle it

Applying the method to tasks whose tool catalogs differ substantially from those encountered during evolution and finding no feasibility gains beyond the 3% baseline would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 2606.12674 by Ching-Yun Ko, Dhaval Patel, Jianxi Gao, Kushal Raj Bhandari, Ling Yue, Pin-Yu Chen, Shaowu Pan.

Figure 1
Figure 1. Figure 1: Dynamic training and evaluation pipeline for Evoflux. A teacher model first generates and refines workflows through execution feedback, typed edits, and LLM-judge scores, producing traces and preference pairs for student training. At evaluation, Evoflux keeps the base small model in a dynamic repair loop that mutates, executes, scores, and selects workflow candidates, while distilled inference uses a train… view at source ↗
Figure 2
Figure 2. Figure 2: Left) Average initial score and average best score per query on the search split. Right) Token efficiency frontier on the search split This scalar score drives parent selection, growth updates, population pruning, and final candidate selection. We also log the six subdimension scores separately, allowing later analysis to dis￾tinguish between task completion, grounding, tool choice, parameter accuracy, dep… view at source ↗
Figure 3
Figure 3. Figure 3: Score distributions across stages for Llama-3.2-3B, Qwen3.5-4B, SmolLM3-3B, and DeepSeek-R1-Qwen-1.5B on the validation split average final score near 3.9. llama3-70b spends roughly 133k tokens and reaches about 3.35. These planners buy quality with a large inference budget. A smaller and more efficient group appears near the lower-left frontier. Nemotron-3-Nano-4B uses about 20k tokens for a score near 1.… view at source ↗
Figure 4
Figure 4. Figure 4: (Left) Prompt, (Middle) completion token use and (Right) execution success by stage for models. most stable improvement at validation time. ReAct becomes attractive when the planner can sustain long trajectories and the application can tolerate higher variance. The next subsection shows that this variance also arrives with a substantially larger token cost. More detailed subskill breakdown is provided in A… view at source ↗
Figure 5
Figure 5. Figure 5: SFT dataset diagnostics. The plots summarize the quality and length profile of the supervised fine tuning data. is the higher scoring candidate, while the rejected workflow is a hard negative drawn from lower scor￾ing alternatives for the same task. The DPO stage uses the SFT checkpoint as the reference model, so preference learning adjusts the supervised model toward workflows that score better under exec… view at source ↗
Figure 6
Figure 6. Figure 6: DPO dataset analysis. The plots summarize pairwise preference data constructed from higher and lower scoring workflow candidates. groups with one, three, six, and a few higher edit counts. This pattern suggests that many rejected candidates come directly from lower-scoring alter￾natives rather than deep edit chains. The dataset still provides useful contrast, but the skew should be treated as a limitation … view at source ↗
Figure 7
Figure 7. Figure 7: Subskill radar plots before and after search on the search split [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Average workflow depth and width across stages. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative validation workflows for Llama-3.2-3B under four regimes. Evoflux produces stronger [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
read the original abstract

Compact language models (LMs) reduce cost, latency, and deployment risk for tool agents. Yet MCP-style tool use requires more than isolated function calling: an agent must discover tools from live catalogs, satisfy schemas, preserve dependencies across intermediate outputs, and ground final responses in executed evidence. Small planners often generate plausible workflow graphs that fail under tool resolution, parameter validation, dependency tracking, or execution. We argue that this failure mode is poorly handled by small-corpus distillation. A few hundred teacher traces can teach workflow format, but rarely cover the recovery behavior needed to repair failed plans over changing tool catalogs. We introduce Evoflux, an inference-time evolutionary search method that treats compact tool use as the repair of executable tool workflows. It evolves typed workflow graphs through structured edits, execution feedback, adaptive intensity, meta-guided redesign, and diversity pruning. On held-out MCP-Bench tasks spanning live MCP servers and 250 tools, Evoflux raises execution feasibility from roughly 3% to 17-24% across small planners. In contrast, SFT and SFT+DPO on the same search-mined data match, underperform, or collapse below zero-shot performance; ReAct reaches higher peaks, but with higher variance and token cost. These results show that execution-grounded search is more reliable under scarce teacher-trace budgets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Evoflux, an inference-time evolutionary search method that evolves typed workflow graphs for tool use by compact language models via structured edits, execution feedback, adaptive intensity, meta-guided redesign, and diversity pruning. It claims that on held-out MCP-Bench tasks with live MCP servers and 250 tools, Evoflux raises execution feasibility from roughly 3% to 17-24% across small planners, outperforming or matching SFT/SFT+DPO on the same mined data while being more reliable than ReAct under scarce teacher-trace budgets, and argues this addresses recovery needs over changing tool catalogs better than distillation.

Significance. If the results hold under rigorous verification, the work would demonstrate a practical inference-time alternative to extensive supervised fine-tuning for improving small models' handling of complex, executable tool workflows, potentially reducing reliance on large teacher traces in dynamic settings.

major comments (2)
  1. [Abstract] Abstract: The central claim that execution feedback plus structured edits and diversity pruning suffice to discover repairs generalizing over changing tool catalogs (without extensive additional teacher traces) is load-bearing for the paper's motivation and conclusions, yet the reported experiments use held-out tasks on fixed live MCP servers with no described catalog-shift protocol (e.g., adding/removing tools between search and evaluation) or measurement of performance degradation under such shifts; the SFT/SFT+DPO baselines therefore cannot isolate whether the evolutionary advantage persists under catalog change.
  2. [Abstract] Abstract: Performance numbers (3% to 17-24% feasibility) are stated without error bars, number of runs, statistical tests, dataset split details, or ablation studies isolating components such as adaptive intensity or meta-guided redesign, preventing verification of the data-to-claim link for the reported improvements over baselines.
minor comments (1)
  1. [Abstract] Abstract: The term 'execution feasibility' is used without a precise definition or reference to how it is measured (e.g., success criteria for workflow resolution, validation, and grounding).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract claims and experimental reporting. We address each major comment below and will revise the manuscript to improve clarity and rigor where the points identify gaps in the current presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that execution feedback plus structured edits and diversity pruning suffice to discover repairs generalizing over changing tool catalogs (without extensive additional teacher traces) is load-bearing for the paper's motivation and conclusions, yet the reported experiments use held-out tasks on fixed live MCP servers with no described catalog-shift protocol (e.g., adding/removing tools between search and evaluation) or measurement of performance degradation under such shifts; the SFT/SFT+DPO baselines therefore cannot isolate whether the evolutionary advantage persists under catalog change.

    Authors: The current experiments evaluate on held-out tasks drawn from the full set of 250 tools on fixed live servers, which tests generalization to unseen task-tool combinations without new teacher traces. The method is designed for inference-time repair using execution feedback, which in principle supports adaptation to catalog variations. However, we agree that an explicit catalog-shift protocol (e.g., tool addition/removal between search and evaluation) is not described and would more directly isolate the advantage over SFT baselines under dynamic conditions. In revision we will clarify the experimental scope in the abstract and main text, add a limitations discussion on this point, and include a targeted catalog-shift analysis if feasible within the page limits. revision: partial

  2. Referee: [Abstract] Abstract: Performance numbers (3% to 17-24% feasibility) are stated without error bars, number of runs, statistical tests, dataset split details, or ablation studies isolating components such as adaptive intensity or meta-guided redesign, preventing verification of the data-to-claim link for the reported improvements over baselines.

    Authors: We agree that the reported feasibility ranges should be accompanied by statistical details for verifiability. The manuscript will be revised to report results over multiple independent runs with error bars or standard deviations, specify the number of runs and any statistical tests used, provide explicit dataset split details, and include ablation studies isolating the contributions of adaptive intensity, meta-guided redesign, and diversity pruning. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method with external benchmarks

full rationale

The paper describes an inference-time evolutionary search procedure (structured edits, execution feedback, diversity pruning) and reports empirical gains on held-out MCP-Bench tasks against baselines (SFT, SFT+DPO, ReAct). No equations, fitted parameters renamed as predictions, self-citations used as load-bearing uniqueness theorems, or ansatzes smuggled via prior work appear in the provided text. The central feasibility claim is grounded in direct experimental comparison on fixed live servers rather than any self-referential reduction of outputs to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5793 in / 1089 out tokens · 27322 ms · 2026-06-27T09:45:26.207461+00:00 · methodology

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 · 16 canonical work pages · 2 internal anchors

  1. [1]

    Belcak, Peter and Heinrich, Greg and Diao, Shizhe and Fu, Yonggan and Dong, Xin and Muralidharan, Saurav and Lin, Yingyan Celine and Molchanov, Pavlo , year = 2025, month = sep, number =. Small. doi:10.48550/arXiv.2506.02153 , urldate =. arXiv , keywords =:2506.02153 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.02153 2025

  2. [2]

    doi:10.48550/arXiv.2602.20133 , urldate =

    Cemri, Mert and Agrawal, Shubham and Gupta, Akshat and Liu, Shu and Cheng, Audrey and Mang, Qiuyang and Naren, Ashwin and Erdogan, Lutfi Eren and Sen, Koushik and Zaharia, Matei and Dimakis, Alex and Stoica, Ion , year = 2026, month = feb, number =. doi:10.48550/arXiv.2602.20133 , urldate =. arXiv , keywords =:2602.20133 , primaryclass =

    work page doi:10.48550/arxiv.2602.20133 2026

  3. [3]

    How Abilities in Large Language Models Are Affected by Supervised Fine-Tuning Data Composition , booktitle =

    Dong, Guanting and Yuan, Hongyi and Lu, Keming and Li, Chengpeng and Xue, Mingfeng and Liu, Dayiheng and Wang, Wei and Yuan, Zheng and Zhou, Chang and Zhou, Jingren , editor =. How Abilities in Large Language Models Are Affected by Supervised Fine-Tuning Data Composition , booktitle =. doi:10.18653/v1/2024.acl-long.12 , abstract =

    work page doi:10.18653/v1/2024.acl-long.12 2024

  4. [4]

    Proceedings of the 2024

    Erdogan, Lutfi Eren and Lee, Nicholas and Jha, Siddharth and Kim, Sehoon and Tabrizi, Ryan and Moon, Suhong and Hooper, Coleman Richard Charles and Anumanchipalli, Gopala and Keutzer, Kurt and Gholami, Amir , editor =. Proceedings of the 2024. doi:10.18653/v1/2024.emnlp-demo.9 , urldate =

    work page doi:10.18653/v1/2024.emnlp-demo.9 2024

  5. [5]

    G., Zhang, T., Wang, X., and Gonzalez, J

    Gonzalez, Joseph and Patil, Shishir and Wang, Xin and Zhang, Tianjun , year = 2024, pages =. Gorilla:. Advances in. doi:10.52202/079017-4020 , urldate =

    work page doi:10.52202/079017-4020 2024

  6. [6]

    International Conference on Learning Representations , author =

  7. [7]

    The Thirty-Ninth Annual Conference on Neural Information Processing Systems , author =

    Distilling. The Thirty-Ninth Annual Conference on Neural Information Processing Systems , author =

  8. [8]

    Luo, Yun and Yang, Zhen and Meng, Fandong and Li, Yafu and Zhou, Jie and Zhang, Yue , year = 2025, journal =. An. doi:10.1109/TASLPRO.2025.3606231 , urldate =

    work page doi:10.1109/taslpro.2025.3606231 2025

  9. [9]

    InFindings of the Association for Computational Linguistics: NAACL 2025, Albuquerque, New Mexico, USA, April 29 - May 4, 2025, Luis Chiruzzo, Alan Ritter, and Lu Wang (Eds.)

    Lu, Jiarui and Holleis, Thomas and Zhang, Yizhe and Aumayer, Bernhard and Nan, Feng and Bai, Haoping and Ma, Shuang and Ma, Shen and Li, Mengyu and Yin, Guoli and Wang, Zirui and Pang, Ruoming , editor =. Findings of the Association for Computational Linguistics:. doi:10.18653/v1/2025.findings-naacl.65 , abstract =

    work page doi:10.18653/v1/2025.findings-naacl.65 2025

  10. [10]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    Novikov, Alexander and V. doi:10.48550/arXiv.2506.13131 , urldate =. arXiv , keywords =:2506.13131 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.13131

  11. [11]

    Advances in Neural Information Processing Systems , author =

    The Berkeley Function Calling Leaderboard (. Advances in Neural Information Processing Systems , author =

  12. [12]

    Qiao, Shuofei and Gui, Honghao and Lv, Chengfei and Jia, Qianghuai and Chen, Huajun and Zhang, Ningyu , editor =. Making. Proceedings of the 2024. doi:10.18653/v1/2024.naacl-long.195 , urldate =

    work page doi:10.18653/v1/2024.naacl-long.195 2024

  13. [13]

    The Twelfth International Conference on Learning Representations , author =

  14. [14]

    and Finn, Chelsea , year = 2023, month = dec, series =

    Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Ermon, Stefano and Manning, Christopher D. and Finn, Chelsea , year = 2023, month = dec, series =. Direct Preference Optimization: Your Language Model Is Secretly a Reward Model , shorttitle =. Proceedings of the 37th

    2023

  15. [15]

    Pawan Kumar, Emilien Dupont, Francisco J

    Mathematical Discoveries from Program Search with Large Language Models , author =. Nature , volume =. doi:10.1038/s41586-023-06924-6 , urldate =

    work page doi:10.1038/s41586-023-06924-6

  16. [16]

    Reflexion: Language Agents with Verbal Reinforcement Learning , booktitle =

  17. [17]

    The Thirty-Ninth Annual Conference on Neural Information Processing Systems , author =

  18. [18]

    The Thirteenth International Conference on Learning Representations , author =

    Scaling. The Thirteenth International Conference on Learning Representations , author =

  19. [19]

    Trial and Error: Exploration-Based Trajectory Optimization of LLM Agents

    Song, Yifan and Yin, Da and Yue, Xiang and Huang, Jie and Li, Sujian and Lin, Bill Yuchen , editor =. Trial and Error:. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1:. doi:10.18653/v1/2024.acl-long.409 , abstract =

    work page doi:10.18653/v1/2024.acl-long.409 2024

  20. [20]

    arXiv preprint arXiv:2508.20453

    Wang, Zhenting and Chang, Qi and Patel, Hemani and Biju, Shashank and Wu, Cheng-En and Liu, Quan and Ding, Aolin and Rezazadeh, Alireza and Shah, Ankit and Bao, Yujia and Siow, Eugene , year = 2025, publisher =. doi:10.48550/ARXIV.2508.20453 , urldate =

    work page doi:10.48550/arxiv.2508.20453 2025

  21. [21]

    Thirty-Seventh Conference on Neural Information Processing Systems , author =

    Tree of Thoughts:. Thirty-Seventh Conference on Neural Information Processing Systems , author =

  22. [22]

    doi:10.18653/v1/2024.emnlp-main.637 , abstract =

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , author =. doi:10.18653/v1/2024.emnlp-main.637 , abstract =

    work page doi:10.18653/v1/2024.emnlp-main.637 2024

  23. [23]

    Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models , booktitle =

    Zhou, Andy and Yan, Kai and. Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models , booktitle =

  24. [24]

    Fireact: Toward language agent fine-tuning.arXiv, 2310.05915, 2023

    Chen, Baian and Shu, Chang and Shareghi, Ehsan and Collier, Nigel and Narasimhan, Karthik and Yao, Shunyu , year = 2023, month = oct, number =. doi:10.48550/arXiv.2310.05915 , urldate =. arXiv , keywords =:2310.05915 , primaryclass =

    work page doi:10.48550/arxiv.2310.05915 2023

  25. [25]

    Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics:

    Zhang, Jianguo and Lan, Tian and Zhu, Ming and Liu, Zuxin and Hoang, Thai and Kokane, Shirley and Yao, Weiran and Tan, Juntao and Liu, Zhiwei and Feng, Yihao and Niebles, Juan Carlos and Heinecke, Shelby and Wang, Huan and Savarese, Silvio and Xiong, Caiming , editor =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Assoc...

    work page doi:10.18653/v1/2025.naacl-long.578 2025

  26. [26]

    Chen, Zehui and Liu, Kuikun and Wang, Qiuchen and Zhang, Wenwei and Liu, Jiangning and Lin, Dahua and Chen, Kai and Zhao, Feng , editor =. Agent-. Findings of the Association for Computational Linguistics:. doi:10.18653/v1/2024.findings-acl.557 , abstract =

    work page doi:10.18653/v1/2024.findings-acl.557 2024

  27. [27]

    The Eleventh International Conference on Learning Representations , author =

  28. [28]

    The Fourteenth International Conference on Learning Representations , author =

  29. [29]

    arXiv preprint arXiv:2603.22386 , year=

    From static templates to dynamic runtime graphs: a survey of workflow optimization for llm agents , author=. arXiv preprint arXiv:2603.22386 , year=

    arXiv

  30. [30]

    2026 , doi =

    FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills , author =. 2026 , doi =

    2026