ARM: Discovering Agentic Reasoning Modules for Generalizable Multi-Agent Systems

Bohan Yao; Shiva Krishna Reddy Malay; Vikas Yadav

arxiv: 2510.05746 · v2 · pith:LI45W7UGnew · submitted 2025-10-07 · 💻 cs.AI · cs.CL· cs.LG

ARM: Discovering Agentic Reasoning Modules for Generalizable Multi-Agent Systems

Bohan Yao , Shiva Krishna Reddy Malay , Vikas Yadav This is my paper

Pith reviewed 2026-05-21 20:43 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.LG

keywords multi-agent systemschain of thoughtagentic reasoningautomatic architecture designgeneralizationtree searchLLM reasoning modules

0 comments

The pith

Agentic Reasoning Modules evolved from Chain of Thought via tree search create multi-agent systems that generalize across models and tasks without retraining.

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

The paper argues that the core limitation in multi-agent systems is not the overall architecture but the quality of the basic reasoning step, and that this step can be automatically improved. It presents a method to evolve a simple Chain of Thought into a more capable Agentic Reasoning Module by searching code variations and using execution-trace reflection to guide mutations. These modules then serve as reusable building blocks that can run recursively or inside a meta-orchestrator. A sympathetic reader would care because the approach promises to replace repeated expensive redesigns for each new task or model with a single discovery process that yields broadly applicable reasoning units.

Core claim

The paper claims that starting from a basic Chain of Thought module and applying tree search over code space with mutations informed by reflection on execution traces reliably discovers Agentic Reasoning Modules. These modules act as versatile reasoning building blocks that can be used directly in recursive loops or as subroutines in a learned meta-orchestrator. Systems assembled from them outperform both hand-designed multi-agent systems and prior automatic design methods while maintaining high performance when transferred to different foundation models and task domains without any further optimization.

What carries the argument

The Agentic Reasoning Module (ARM), an evolved specialization of Chain of Thought in which each granular reasoning step is handled by a code module discovered through reflection-guided tree search over the code space.

If this is right

Multi-agent systems assembled from ARM modules achieve higher accuracy on complex reasoning tasks than both manual designs and earlier automatic methods.
The same discovered modules maintain strong performance when the underlying foundation model is swapped.
No re-discovery of architectures or new labeled validation data is required when moving to a different task domain.
ARM can be deployed either as a standalone recursive loop or as a subroutine inside a learned meta-orchestrator.

Where Pith is reading between the lines

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

The evolutionary process may uncover reusable patterns of reasoning that transfer to entirely new problem classes beyond those tested.
If the modules prove stable, future systems could be built by composing a small library of such discovered units rather than designing from scratch each time.
The reflection step in the search could be adapted to discover modules specialized for planning or tool selection in addition to reasoning.
Widespread adoption would lower the barrier for deploying reliable agentic systems in settings where labeled data or compute for repeated searches is unavailable.

Load-bearing premise

That mutations guided by reflection on execution traces, applied through tree search starting from simple Chain of Thought, will produce reasoning modules that are both more capable and more generalizable than standard Chain of Thought or existing multi-agent designs.

What would settle it

An experiment in which multi-agent systems built from the discovered modules show no performance gain over plain Chain of Thought or lose accuracy when tested on a new task domain or different foundation model without additional tuning.

Figures

Figures reproduced from arXiv: 2510.05746 by Bohan Yao, Shiva Krishna Reddy Malay, Vikas Yadav.

**Figure 1.** Figure 1: An illustration of the proposed ARM module on the left and the meta policy on the right using "Self refine" as an example MAS. The ARM module takes a question and previous reasoning steps and executes a MAS to get the next step. The meta policy uses ARM as a sub-module and orchestrates the overarching global strategy. Note that this is for illustration only, the actual step generator and the meta policy di… view at source ↗

**Figure 2.** Figure 2: Validation of the meta-policy transfer for top discovered policies. The table compares performance using the simple surrogate mCoT (CoT Baseline) versus the powerful ARM module m∗ (Meta Policy). The intermediate CoT→Meta column isolates the performance gain from the superior m∗ module alone by evaluating it on states generated by the baseline [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

read the original abstract

Large Language Model (LLM)-powered Multi-agent systems (MAS) have achieved state-of-the-art results on various complex reasoning tasks. Recent works have proposed techniques to automate the design of MASes, eliminating the need for manual engineering. However, these techniques perform poorly, often achieving similar or inferior performance to simple baselines. Furthermore, they require computationally expensive re-discovery of architectures for each new task domain and expensive data annotation on domains without existing labeled validation sets. A critical insight is that simple Chain of Thought (CoT) reasoning often performs competitively with these complex systems, suggesting that the fundamental reasoning unit of MASes, CoT, warrants further investigation. To this end, we present a new paradigm for automatic MAS design that pivots the focus to optimizing CoT reasoning. We introduce the Agentic Reasoning Module (ARM), an agentic generalization of CoT where each granular reasoning step is executed by a specialized reasoning module. This module is discovered through a tree search over the code space, starting from a simple CoT module and evolved using mutations informed by reflection on execution traces. The resulting ARM acts as a versatile reasoning building block which can be utilized as a direct recursive loop or as a subroutine in a learned meta-orchestrator. Our approach significantly outperforms both manually designed MASes and state-of-the-art automatic MAS design methods. Crucially, MASes built with ARM exhibit superb generalization, maintaining high performance across different foundation models and task domains without further optimization.

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.

Desk Editor's Note private letter to a colleague

The paper's main idea is to evolve reusable reasoning modules from basic CoT via code-space tree search and reflection, aiming for MAS that transfer across models and domains without re-optimization.

read the letter

The paper's main idea is to evolve reusable reasoning modules from basic CoT via code-space tree search and reflection, aiming for MAS that transfer across models and domains without re-optimization. This shifts the target from whole architectures to the reasoning primitive itself, which is a distinct framing from most prior automatic MAS design work that searches over agent roles or connections instead. Starting from a simple CoT module and using execution-trace reflection to guide mutations in code space could produce building blocks that serve as recursive loops or subroutines in a meta-orchestrator. That focus on the fundamental unit makes sense given the abstract's point that plain CoT already competes with complex systems. The approach also tries to avoid expensive per-domain re-discovery and annotation, which is a practical pain point in the area. If the search reliably yields more capable and transferable modules, it would reduce engineering overhead when swapping models or tasks. The description of the process is clear enough on paper to see how it differs from standard search methods. The generalization claim is the part that needs the most scrutiny. The abstract says MASes built with ARM maintain high performance across different foundation models and task domains without further optimization, but it gives no numbers, baselines, or details on how the transfer was tested. Reflection is done by an LLM, so the discovered modules could embed patterns tied to the search-time model rather than truly model-agnostic structures. If the experiments only use similar model families or related domains, the transfer result would not fully support the claim. The tree search cost itself is not addressed in the abstract either. This work would interest researchers focused on making multi-agent systems more modular and less brittle to changes in models or tasks. A reader looking for concrete methods to improve reasoning steps without heavy manual design would find the core proposal worth examining. I would send it for peer review. The angle on optimizing the CoT unit through code search is different enough from existing literature that referees should evaluate the full experiments and controls.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the Agentic Reasoning Module (ARM) as an agentic generalization of Chain-of-Thought (CoT) reasoning. ARM modules are discovered via tree search over code space, starting from a simple CoT module and evolved through mutations guided by reflection on execution traces. The authors claim that MASes constructed with these ARM modules significantly outperform both manually designed MASes and prior automatic MAS design methods, while also exhibiting superb generalization across different foundation models and task domains without any further optimization or re-discovery.

Significance. If the generalization and performance claims are substantiated, the work would represent a meaningful shift in automatic MAS design by redirecting effort from architecture search to optimization of the core reasoning unit. This could reduce the computational cost of per-domain re-discovery highlighted in the abstract and provide a reusable building block for both recursive loops and meta-orchestrators.

major comments (3)

[§4.2] §4.2 (Generalization Experiments): The cross-model and cross-domain results are central to the 'superb generalization without further optimization' claim, yet the manuscript does not report whether the reflection LLM used during module discovery is held fixed or matches the test-time foundation models; this leaves open the possibility that discovered modules embed model-specific patterns rather than domain- and model-agnostic structures.
[Table 3] Table 3 (Cross-domain transfer results): The reported performance maintenance across domains lacks error bars, number of runs, and statistical significance tests; without these, it is difficult to assess whether the observed generalization is robust or could be explained by variance in the underlying LLMs.
[§3.1] §3.1 (Mutation and Reflection Mechanism): The description of how execution-trace reflection informs code mutations is high-level; a concrete example of a mutation step (input trace, reflection output, resulting code change) would be required to evaluate whether the search reliably produces more generalizable modules than standard CoT.

minor comments (2)

[Figure 2] Figure 2 (ARM architecture diagram): The recursive loop versus meta-orchestrator usage modes are not visually distinguished, making it hard to follow how a single discovered module is deployed in both settings.
[§5] The abstract states that simple CoT 'often performs competitively' with complex MASes, but the experimental section does not include a direct head-to-head comparison of ARM against an optimized single CoT baseline under the same generalization protocol.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below and have revised the manuscript accordingly to improve clarity and rigor.

read point-by-point responses

Referee: [§4.2] §4.2 (Generalization Experiments): The cross-model and cross-domain results are central to the 'superb generalization without further optimization' claim, yet the manuscript does not report whether the reflection LLM used during module discovery is held fixed or matches the test-time foundation models; this leaves open the possibility that discovered modules embed model-specific patterns rather than domain- and model-agnostic structures.

Authors: We clarify that the reflection LLM is held fixed throughout module discovery (specifically, we use GPT-4o for all reflection steps in the tree search). The discovered ARM modules are then deployed at test time with different foundation models without any re-optimization or re-discovery. This separation is central to our generalization claim. We will explicitly state this fixed-reflection setup and its implications in the revised §4.2. revision: yes
Referee: [Table 3] Table 3 (Cross-domain transfer results): The reported performance maintenance across domains lacks error bars, number of runs, and statistical significance tests; without these, it is difficult to assess whether the observed generalization is robust or could be explained by variance in the underlying LLMs.

Authors: We agree that additional statistical detail is needed. In the revision we will report results over 5 independent runs with error bars (standard deviation), explicitly state the number of runs, and include paired t-test p-values comparing ARM-based MASes against baselines to confirm that the cross-domain improvements are statistically significant. revision: yes
Referee: [§3.1] §3.1 (Mutation and Reflection Mechanism): The description of how execution-trace reflection informs code mutations is high-level; a concrete example of a mutation step (input trace, reflection output, resulting code change) would be required to evaluate whether the search reliably produces more generalizable modules than standard CoT.

Authors: We will add a concrete worked example to §3.1 (or a new appendix figure) showing an input execution trace, the exact reflection output produced by the LLM, and the resulting code mutation. This example will illustrate how reflection identifies a specific inefficiency and how the mutation improves generality beyond basic CoT. revision: yes

Circularity Check

0 steps flagged

No significant circularity; generalization claim is empirical

full rationale

The paper describes an empirical discovery process: starting from a simple CoT module, performing tree search over code space with mutations guided by reflection on execution traces, then using the resulting ARM as a building block. The central claim of superb generalization across foundation models and task domains without further optimization is presented as an observed experimental outcome rather than a mathematical derivation or fitted prediction that reduces to the search objective by construction. No equations, self-definitional loops, or load-bearing self-citations are invoked in the provided text to force the result. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central claim rests on the untested premise that evolutionary search over code will discover modules whose generalization properties transfer across models and domains. No free parameters, axioms, or invented entities are explicitly quantified in the abstract.

invented entities (1)

Agentic Reasoning Module (ARM) no independent evidence
purpose: A specialized, evolvable code module that replaces a single CoT step and can be reused recursively or inside a meta-orchestrator.
Introduced as the core new building block; no independent evidence of its existence outside the search process is provided in the abstract.

pith-pipeline@v0.9.0 · 5799 in / 1335 out tokens · 31446 ms · 2026-05-21T20:43:10.132521+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We discover both the optimal step-generator m* and meta-policy π* using a unified Reflection-Guided Evolutionary Search algorithm. This algorithm performs a tree search over the programmatic space of valid Python modules...
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the scaffolded surrogate objective... evaluates it within the stable context of a reference trace generated by the baseline m_CoT

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 16 internal anchors

[1]

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Ar- nav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alexandros G Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, and Omar Khattab. GEPA: Reflec- tive prompt evolution can outperform reinforcement learning.arXiv preprint arXi...

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Fireact: Toward language agent fine-tuning.arXiv preprint arXiv:2403.01925,

Baian Chen, Chang Li, Zhuo Li, Jianing Wang, Yapen Tian, Rui Wang, and Xin Wang. Fireact: Toward language agent fine-tuning.arXiv preprint arXiv:2403.01925,

work page arXiv
[3]

AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors

Weize Chen, Yusheng Zhang, Zihan Zhang, Cheng Liu, Zipeng Zheng, Chen Qian, Yufan Zhao, Yufan Cong, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors.arXiv preprint arXiv:2308.10848,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

16 Solving math word problems with process- and outcome-based feedback A

Antonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large lan- guage models for interpretable logical reasoning.arXiv preprint arXiv:2205.09712,

work page arXiv
[5]

Self-collaboration code generation via chatgpt

Yihong Dong, Xue Wang, Ge Jiang, Zhiping Liu, Cilin Zhang, Peiyu Wang, and Yi Zhang. Self- collaboration code generation via chatgpt.arXiv preprint arXiv:2304.07590,

work page arXiv
[6]

Improving Factuality and Reasoning in Language Models through Multiagent Debate

URLhttps: //arxiv.org/abs/2305.14325. Rishabh Dubey, Daochen Zha, Lingjiao Wu, and Aditya Grover. Revisiting the gold standard: A critical look at multi-agent systems for discovery.arXiv preprint arXiv:2310.10653,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Flowreasoner: Reinforcing query-level meta-agents.arXiv preprint arXiv:2504.15257, 2025

URLhttps://proceedings.mlr.press/v235/fernando24a.html. Hongcheng Gao, Yue Liu, Yufei He, Longxu Dou, Chao Du, Zhijie Deng, Bryan Hooi, Min Lin, and Tianyu Pang. Flowreasoner: Reinforcing query-level meta-agents.arXiv preprint arXiv:2504.15257,

work page arXiv
[8]

Reasoning with Language Model is Planning with World Model

Shibo Hao, Yi Gu, Haodi Ma, Joshua Wang, Zhen Chen, and Zhaofeng Wang. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

Sirui Hong, Xiawu Zheng, Jonathan Chen, Kechen Yang, Yida Li, Weya Su, Chen Wang, Ceyao He, et al. Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

https://arxiv.org/abs/2504.09037

Morgan Kaufmann Publishers Inc. ISBN 1558608737. 11 Preprint Zixuan Ke, Fangkai Jiao, Yifei Ming, Xuan-Phi Nguyen, Austin Xu, Do Xuan Long, Minzhi Li, Chengwei Qin, Peifeng Wang, Silvio Savarese, et al. A survey of frontiers in llm reasoning: Inference scaling, learning to reason, and agentic systems.arXiv preprint arXiv:2504.09037,

work page arXiv
[11]

Aflow: Automating agentic workflow generation for large language models

Sungwoo Kim, Lin Xu, Yifan Guo, Arif Rahman, and Shiyi Wang. Aflow: Automating agentic workflow generation for large language models. InProceedings of the 2024 Conference on Em- pirical Methods in Natural Language Processing (EMNLP),

work page 2024
[12]

Large Language Models are Zero-Shot Reasoners

Accessed: YYYY-MM-DD. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.arXiv preprint arXiv:2205.11916,

work page internal anchor Pith review Pith/arXiv arXiv
[13]

Guohao Li, Hasan Momin, Hasan Ground, ‘ Kian, et al

URLhttps://arxiv.org/html/2505.21963v1. Guohao Li, Hasan Momin, Hasan Ground, ‘ Kian, et al. Camel: Communicative agents for" mind" exploration of large scale language model society.arXiv preprint arXiv:2303.17760,

work page arXiv
[14]

Self-Refine: Iterative Refinement with Self-Feedback

Aman Madaan, Niket Tandon, Prakhar Gupta, Kevin Hall, Luyu Gao, Sarah Wiegreffe, Uri Alon, Pengcheng Cair, et al. Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

AlphaEvolve: A coding agent for scientific and algorithmic discovery

Accessed: 2025-09-27. Alexander Novikov, Ngân V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent f...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[16]

Show Your Work: Scratchpads for Intermediate Computation with Language Models

Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Noah Chapman, George Dugan, Miljan Tworkowski, Croitoru Alfredo, et al. Show your work: Scratchpads for intermediate computation with language models.arXiv preprint arXiv:2112.00114,

work page internal anchor Pith review Pith/arXiv arXiv
[17]

Introducing gpt-4.1 in the api.https://openai.com/index/gpt-4-1/, April 2025a

OpenAI. Introducing gpt-4.1 in the api.https://openai.com/index/gpt-4-1/, April 2025a. Accessed: 2025-09-27. OpenAI. Introducing openai o3 and o4-mini.https://openai.com/index/ introducing-o3-and-o4-mini/, April 2025b. Accessed: 2025-09-27. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Wel...

work page 2025
[18]

GPT-4o System Card

URLhttps://arxiv.org/abs/2410.21276. Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior.arXiv preprint arXiv:2304.03442,

work page internal anchor Pith review Pith/arXiv arXiv
[19]

ChatDev: Communicative Agents for Software Development

13 Preprint Chen Qian, Xin Wang, Yufan Cong, Cheng Liu, Weize Yu, Zipeng Zheng, Zihan Chen, Yapen Gu, et al. Communicative agents for software development.arXiv preprint arXiv:2307.07924,

work page internal anchor Pith review Pith/arXiv arXiv
[20]

GPQA: A Graduate-Level Google-Proof Q&A Benchmark

David Rein, Ansh Raichur, Caleb Riddoch, Andrew Andreassen, Ben Jones, Zihui Wu, Shufan Jiang, Kevin Chen, Cong Jiang, Andy Zhao, Lucy Yuan, Jerry Li, Yaofeng Zhang, R Ar- jun Gopalakrishnan, Andrew Pan, Yapei Zhou, Leon Tang, Thomas Lee, Tom Brown, and Jacob Steinhardt. GPQA: A graduate-level google-proof q&a benchmark.arXiv preprint arXiv:2311.12022,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Reflexion: Language Agents with Verbal Reinforcement Learning

URLhttps://blog.ml.cmu.edu/2024/12/06/ scribeagent-fine-tuning-open-source-llms-for-enhanced-web-navigation/. Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: Language agents with verbal rein- forcement learning.arXiv preprint arXiv:2303.11366,

work page internal anchor Pith review Pith/arXiv arXiv 2024
[22]

Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song

ISBN 0262039249. Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song. Rethinking the bounds of llm reasoning: Are multi-agent discussions the key? InarXiv preprint arXiv:2402.18272,

work page arXiv
[23]

Ruida Wang, Rui Pan, Yuxin Li, Jipeng Zhang, Yizhen Jia, Shizhe Diao, Renjie Pi, Junjie Hu, and Tong Zhang

Accessed: YYYY-MM-DD. Ruida Wang, Rui Pan, Yuxin Li, Jipeng Zhang, Yizhen Jia, Shizhe Diao, Renjie Pi, Junjie Hu, and Tong Zhang. Ma-lot: Multi-agent lean-based long chain-of-thought reasoning enhances formal theorem proving.arXiv preprint arXiv:2503.03205,

work page arXiv
[24]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

work page internal anchor Pith review Pith/arXiv arXiv
[25]

AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Li, Erkang Zhu, Beibin Li, Li Jiang, et al. Autogen: Enabling next-gen llm applications via multi-agent conversation.arXiv preprint arXiv:2308.08155,

work page internal anchor Pith review Pith/arXiv arXiv
[26]

A toolbox, not a hammer–multi-tag: Scaling math reasoning with multi-tool aggregation.arXiv preprint arXiv:2507.18973,

Bohan Yao and Vikas Yadav. A toolbox, not a hammer–multi-tag: Scaling math reasoning with multi-tool aggregation.arXiv preprint arXiv:2507.18973,

work page arXiv
[27]

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Sha, Silvio Savarese, and Sima an. Tree of thoughts: Deliberate problem solving with large language models.arXiv preprint arXiv:2305.10601, 2023a. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Sha, Narasimhan Karthik, and Sima an. React: Synergizing reasoning and acting in language models. InThe Eleventh Interna...

work page internal anchor Pith review Pith/arXiv arXiv
[28]

Flowreasoner: Automatic multi- agent system generation for complex reasoning.arXiv preprint arXiv:2502.08123, 2025c

Wei Zhang, Chen Liu, Ananya Patel, Ming Zhao, and Jie Huang. Flowreasoner: Automatic multi- agent system generation for complex reasoning.arXiv preprint arXiv:2502.08123, 2025c. Ac- cessed: YYYY-MM-DD. Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engine...

work page arXiv
[29]

15 Preprint A THEORETICALANALYSIS A complete theoretical analysis of the multi-agentic system ARM powered by LLMs is intractable due to the complex, high-dimensional nature of language generation and the non-stationary of the generation process. Therefore, to build a formal intuition for the design choices in our scaffolded search for the step-generator, ...

work page 2018
[30]

baseline–candidate–baseline

unconstrained search space of code modules. 16 Preprint A.2 THEORETICALGROUNDING FOR THESCAFFOLDEDSTEP-GENERATORSEARCH The scaffolded objective evaluates a candidatembysplicingit into a baseline rollout for a short windowt∈ {i, . . . , i+ℓ−1}while keepingm CoT before and after: U ∗ mCoT ◦ U ℓ m ◦U i mCoT | {z } “baseline–candidate–baseline” . Letd CoT,t b...

work page 2002

[1] [1]

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Ar- nav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alexandros G Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, and Omar Khattab. GEPA: Reflec- tive prompt evolution can outperform reinforcement learning.arXiv preprint arXi...

work page internal anchor Pith review Pith/arXiv arXiv

[2] [2]

Fireact: Toward language agent fine-tuning.arXiv preprint arXiv:2403.01925,

Baian Chen, Chang Li, Zhuo Li, Jianing Wang, Yapen Tian, Rui Wang, and Xin Wang. Fireact: Toward language agent fine-tuning.arXiv preprint arXiv:2403.01925,

work page arXiv

[3] [3]

AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors

Weize Chen, Yusheng Zhang, Zihan Zhang, Cheng Liu, Zipeng Zheng, Chen Qian, Yufan Zhao, Yufan Cong, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors.arXiv preprint arXiv:2308.10848,

work page internal anchor Pith review Pith/arXiv arXiv

[4] [4]

16 Solving math word problems with process- and outcome-based feedback A

Antonia Creswell, Murray Shanahan, and Irina Higgins. Selection-inference: Exploiting large lan- guage models for interpretable logical reasoning.arXiv preprint arXiv:2205.09712,

work page arXiv

[5] [5]

Self-collaboration code generation via chatgpt

Yihong Dong, Xue Wang, Ge Jiang, Zhiping Liu, Cilin Zhang, Peiyu Wang, and Yi Zhang. Self- collaboration code generation via chatgpt.arXiv preprint arXiv:2304.07590,

work page arXiv

[6] [6]

Improving Factuality and Reasoning in Language Models through Multiagent Debate

URLhttps: //arxiv.org/abs/2305.14325. Rishabh Dubey, Daochen Zha, Lingjiao Wu, and Aditya Grover. Revisiting the gold standard: A critical look at multi-agent systems for discovery.arXiv preprint arXiv:2310.10653,

work page internal anchor Pith review Pith/arXiv arXiv

[7] [7]

Flowreasoner: Reinforcing query-level meta-agents.arXiv preprint arXiv:2504.15257, 2025

URLhttps://proceedings.mlr.press/v235/fernando24a.html. Hongcheng Gao, Yue Liu, Yufei He, Longxu Dou, Chao Du, Zhijie Deng, Bryan Hooi, Min Lin, and Tianyu Pang. Flowreasoner: Reinforcing query-level meta-agents.arXiv preprint arXiv:2504.15257,

work page arXiv

[8] [8]

Reasoning with Language Model is Planning with World Model

Shibo Hao, Yi Gu, Haodi Ma, Joshua Wang, Zhen Chen, and Zhaofeng Wang. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992,

work page internal anchor Pith review Pith/arXiv arXiv

[9] [9]

MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

Sirui Hong, Xiawu Zheng, Jonathan Chen, Kechen Yang, Yida Li, Weya Su, Chen Wang, Ceyao He, et al. Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352,

work page internal anchor Pith review Pith/arXiv arXiv

[10] [10]

https://arxiv.org/abs/2504.09037

Morgan Kaufmann Publishers Inc. ISBN 1558608737. 11 Preprint Zixuan Ke, Fangkai Jiao, Yifei Ming, Xuan-Phi Nguyen, Austin Xu, Do Xuan Long, Minzhi Li, Chengwei Qin, Peifeng Wang, Silvio Savarese, et al. A survey of frontiers in llm reasoning: Inference scaling, learning to reason, and agentic systems.arXiv preprint arXiv:2504.09037,

work page arXiv

[11] [11]

Aflow: Automating agentic workflow generation for large language models

Sungwoo Kim, Lin Xu, Yifan Guo, Arif Rahman, and Shiyi Wang. Aflow: Automating agentic workflow generation for large language models. InProceedings of the 2024 Conference on Em- pirical Methods in Natural Language Processing (EMNLP),

work page 2024

[12] [12]

Large Language Models are Zero-Shot Reasoners

Accessed: YYYY-MM-DD. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners.arXiv preprint arXiv:2205.11916,

work page internal anchor Pith review Pith/arXiv arXiv

[13] [13]

Guohao Li, Hasan Momin, Hasan Ground, ‘ Kian, et al

URLhttps://arxiv.org/html/2505.21963v1. Guohao Li, Hasan Momin, Hasan Ground, ‘ Kian, et al. Camel: Communicative agents for" mind" exploration of large scale language model society.arXiv preprint arXiv:2303.17760,

work page arXiv

[14] [14]

Self-Refine: Iterative Refinement with Self-Feedback

Aman Madaan, Niket Tandon, Prakhar Gupta, Kevin Hall, Luyu Gao, Sarah Wiegreffe, Uri Alon, Pengcheng Cair, et al. Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651,

work page internal anchor Pith review Pith/arXiv arXiv

[15] [15]

AlphaEvolve: A coding agent for scientific and algorithmic discovery

Accessed: 2025-09-27. Alexander Novikov, Ngân V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent f...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[16] [16]

Show Your Work: Scratchpads for Intermediate Computation with Language Models

Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Noah Chapman, George Dugan, Miljan Tworkowski, Croitoru Alfredo, et al. Show your work: Scratchpads for intermediate computation with language models.arXiv preprint arXiv:2112.00114,

work page internal anchor Pith review Pith/arXiv arXiv

[17] [17]

Introducing gpt-4.1 in the api.https://openai.com/index/gpt-4-1/, April 2025a

OpenAI. Introducing gpt-4.1 in the api.https://openai.com/index/gpt-4-1/, April 2025a. Accessed: 2025-09-27. OpenAI. Introducing openai o3 and o4-mini.https://openai.com/index/ introducing-o3-and-o4-mini/, April 2025b. Accessed: 2025-09-27. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Wel...

work page 2025

[18] [18]

GPT-4o System Card

URLhttps://arxiv.org/abs/2410.21276. Joon Sung Park, Joseph C O’Brien, Carrie J Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior.arXiv preprint arXiv:2304.03442,

work page internal anchor Pith review Pith/arXiv arXiv

[19] [19]

ChatDev: Communicative Agents for Software Development

13 Preprint Chen Qian, Xin Wang, Yufan Cong, Cheng Liu, Weize Yu, Zipeng Zheng, Zihan Chen, Yapen Gu, et al. Communicative agents for software development.arXiv preprint arXiv:2307.07924,

work page internal anchor Pith review Pith/arXiv arXiv

[20] [20]

GPQA: A Graduate-Level Google-Proof Q&A Benchmark

David Rein, Ansh Raichur, Caleb Riddoch, Andrew Andreassen, Ben Jones, Zihui Wu, Shufan Jiang, Kevin Chen, Cong Jiang, Andy Zhao, Lucy Yuan, Jerry Li, Yaofeng Zhang, R Ar- jun Gopalakrishnan, Andrew Pan, Yapei Zhou, Leon Tang, Thomas Lee, Tom Brown, and Jacob Steinhardt. GPQA: A graduate-level google-proof q&a benchmark.arXiv preprint arXiv:2311.12022,

work page internal anchor Pith review Pith/arXiv arXiv

[21] [21]

Reflexion: Language Agents with Verbal Reinforcement Learning

URLhttps://blog.ml.cmu.edu/2024/12/06/ scribeagent-fine-tuning-open-source-llms-for-enhanced-web-navigation/. Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: Language agents with verbal rein- forcement learning.arXiv preprint arXiv:2303.11366,

work page internal anchor Pith review Pith/arXiv arXiv 2024

[22] [22]

Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song

ISBN 0262039249. Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song. Rethinking the bounds of llm reasoning: Are multi-agent discussions the key? InarXiv preprint arXiv:2402.18272,

work page arXiv

[23] [23]

Ruida Wang, Rui Pan, Yuxin Li, Jipeng Zhang, Yizhen Jia, Shizhe Diao, Renjie Pi, Junjie Hu, and Tong Zhang

Accessed: YYYY-MM-DD. Ruida Wang, Rui Pan, Yuxin Li, Jipeng Zhang, Yizhen Jia, Shizhe Diao, Renjie Pi, Junjie Hu, and Tong Zhang. Ma-lot: Multi-agent lean-based long chain-of-thought reasoning enhances formal theorem proving.arXiv preprint arXiv:2503.03205,

work page arXiv

[24] [24]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

work page internal anchor Pith review Pith/arXiv arXiv

[25] [25]

AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Li, Erkang Zhu, Beibin Li, Li Jiang, et al. Autogen: Enabling next-gen llm applications via multi-agent conversation.arXiv preprint arXiv:2308.08155,

work page internal anchor Pith review Pith/arXiv arXiv

[26] [26]

A toolbox, not a hammer–multi-tag: Scaling math reasoning with multi-tool aggregation.arXiv preprint arXiv:2507.18973,

Bohan Yao and Vikas Yadav. A toolbox, not a hammer–multi-tag: Scaling math reasoning with multi-tool aggregation.arXiv preprint arXiv:2507.18973,

work page arXiv

[27] [27]

Tree of Thoughts: Deliberate Problem Solving with Large Language Models

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Sha, Silvio Savarese, and Sima an. Tree of thoughts: Deliberate problem solving with large language models.arXiv preprint arXiv:2305.10601, 2023a. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Sha, Narasimhan Karthik, and Sima an. React: Synergizing reasoning and acting in language models. InThe Eleventh Interna...

work page internal anchor Pith review Pith/arXiv arXiv

[28] [28]

Flowreasoner: Automatic multi- agent system generation for complex reasoning.arXiv preprint arXiv:2502.08123, 2025c

Wei Zhang, Chen Liu, Ananya Patel, Ming Zhao, and Jie Huang. Flowreasoner: Automatic multi- agent system generation for complex reasoning.arXiv preprint arXiv:2502.08123, 2025c. Ac- cessed: YYYY-MM-DD. Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engine...

work page arXiv

[29] [29]

15 Preprint A THEORETICALANALYSIS A complete theoretical analysis of the multi-agentic system ARM powered by LLMs is intractable due to the complex, high-dimensional nature of language generation and the non-stationary of the generation process. Therefore, to build a formal intuition for the design choices in our scaffolded search for the step-generator, ...

work page 2018

[30] [30]

baseline–candidate–baseline

unconstrained search space of code modules. 16 Preprint A.2 THEORETICALGROUNDING FOR THESCAFFOLDEDSTEP-GENERATORSEARCH The scaffolded objective evaluates a candidatembysplicingit into a baseline rollout for a short windowt∈ {i, . . . , i+ℓ−1}while keepingm CoT before and after: U ∗ mCoT ◦ U ℓ m ◦U i mCoT | {z } “baseline–candidate–baseline” . Letd CoT,t b...

work page 2002