arxiv: 2605.14483 · v1 · submitted 2026-05-14 · 💻 cs.AI

Recognition: no theorem link

LEMON: Learning Executable Multi-Agent Orchestration via Counterfactual Reinforcement Learning

Xudong Chen , Yixin Liu , Hua Wei , Kaize Ding

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:10 UTC · model grok-4.3

classification 💻 cs.AI

keywords multi-agent orchestrationcounterfactual reinforcement learningLLM agentscredit assignmentexecutable specificationsreasoning benchmarkscoding benchmarksGRPO objective

0 comments

The pith

Training via localized counterfactual edits allows an LLM to generate executable multi-agent orchestrations that outperform prior methods on reasoning and coding benchmarks.

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

The paper aims to show that augmenting standard reinforcement learning objectives with counterfactual signals from editing specific orchestration components provides better training signals for designing multi-agent systems. By isolating changes to roles, capacities, or dependencies and using the resulting reward differences for credit assignment, the method avoids the diffuse feedback of whole-system execution outcomes. This matters because effective multi-agent performance often hinges on precise coordination that is hard to discover through trial and error alone. If the training works as described, it automates the creation of task-adapted agent teams in a way that integrates all design elements at once rather than piecemeal.

Core claim

LEMON is an LLM-based orchestrator trained to output an executable specification that combines task-specific roles, customized duties, capacity levels, and dependency structures. The training augments the GRPO objective with a localized counterfactual signal obtained by editing individual orchestration fields and applying the reward contrast exclusively to the changed spans. This yields state-of-the-art results among compared multi-agent orchestration methods on the MMLU, GSM8K, AQuA, MultiArith, SVAMP, and HumanEval benchmarks.

What carries the argument

Localized counterfactual signal that edits single fields such as role or dependency and feeds the reward contrast only to those edited parts of the orchestration specification.

If this is right

Orchestration design receives more precise credit assignment than whole-run feedback allows.
All elements of the multi-agent system are optimized jointly in one specification.
The output is immediately executable as a deployable system.
Performance advantages appear on both mathematical reasoning and code generation tasks.

Where Pith is reading between the lines

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

Similar localized counterfactual training could be used to refine other LLM-generated plans or workflows beyond agents.
The approach may reduce reliance on human-designed templates by learning from reward contrasts alone.
Testing the method on agent systems with dynamic or runtime-changing dependencies would check its robustness.

Load-bearing premise

Editing single orchestration fields and measuring the resulting reward contrast supplies reliable, localized credit assignment superior to standard execution-level feedback.

What would settle it

An ablation experiment in which the counterfactual editing step is removed and benchmark performance remains unchanged would show that the localized signal is not driving the gains.

Figures

Figures reproduced from arXiv: 2605.14483 by Hua Wei, Kaize Ding, Xudong Chen, Yixin Liu.

**Figure 1.** Figure 1: Comparison of decomposed and compositional orchestration generation under sparse and localized credit assignment. Challenge 1: Decomposed design of executable orchestration specifications. Existing methods often optimize only part of the orchestration design space at the level of individual design components, rather than at the level of the final executable specification, as illustrated by the decompose… view at source ↗

**Figure 2.** Figure 2: Overview of LEMON. The orchestrator generates an executable orchestration specification that composes task-specific roles, capacity levels, and dependency references, and is trained with orchestration-level GRPO and localized counterfactual credit assignment. 2 Related Work Multi-Agent Collaboration and Orchestration Design. Recent work has moved LLM-based multi-agent systems beyond fixed manual designs by… view at source ↗

**Figure 3.** Figure 3: The worker token cost comparison. orchestration adaptation. They also support our central claim that role specifications, capacity levels, and dependency structure should be generated jointly as a single executable specification. Token Efficiency. Beyond task performance, we evaluate the accuracy–token trade-off of different methods [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Case studies on GSM8K and MMLU. Node colors denote capacity levels, and the expanded [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Effect of local counterfactual credit assignment on reward, efficiency, and mutation [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

read the original abstract

Large language models (LLMs) have become a strong foundation for multi-agent systems, but their effectiveness depends heavily on orchestration design. Across different tasks, role design, capacity assignment, and dependency construction jointly affect both solution quality and execution efficiency. Existing approaches automate parts of this design process, yet they often optimize these decisions partially or sequentially, and rely on execution-level feedback that provides limited credit assignment for local orchestration decisions. We propose LEMON (\textbf{L}earning \textbf{E}xecutable \textbf{M}ulti-agent \textbf{O}rchestratio\textbf{N} via Counterfactual Reinforcement Learning), an LLM-based orchestrator that generates an executable orchestration specification. The specification integrates task-specific roles, customized duties, capacity levels, and dependency structure into a single deployable system. To train the orchestrator, we augment the orchestration-level GRPO objective with a localized counterfactual signal that edits role, capacity, or dependency fields and applies the resulting reward contrast only to the edited spans. Experiments on six reasoning and coding benchmarks, including MMLU, GSM8K, AQuA, MultiArith, SVAMP, and HumanEval, show that LEMON achieves state-of-the-art performance among the evaluated multi-agent orchestration methods. Our code is available at https://anonymous.4open.science/r/LEMON-B23C.

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

LEMON adds localized counterfactual edits to GRPO for training multi-agent orchestration specs, but cascading effects likely undermine the claimed credit isolation.

read the letter

Colleague, the main point is that LEMON trains an LLM to output full executable orchestration specs for multi-agent systems by augmenting the GRPO objective with reward contrasts from single-field edits on roles, capacities, or dependencies. It reports better results than prior orchestration methods on MMLU, GSM8K, AQuA, MultiArith, SVAMP, and HumanEval, with code released for the setup. What the paper does well is move past sequential or partial optimization by generating the entire specification at once and tying the training signal more directly to local changes in the output. That addresses a real pain point in LLM multi-agent work where execution feedback gives weak guidance on individual design choices. The approach is straightforward to implement on top of existing RL objectives and the benchmarks are standard ones. The soft spot is the assumption that editing one field produces a reward delta mostly attributable to that field. In practice, altering a dependency or capacity often shifts which agents activate and in what sequence, so the contrast mixes global effects. The abstract gives no ablations that hold the trajectories fixed and compare localized versus full reward application, nor any breakdown of how often edits stay local. Without those checks the central technical advantage stays unproven even if the overall numbers look good. This is for groups already working on automated design of LLM agent teams who need a concrete training recipe. It is worth a serious referee because the problem is timely, the method is specified enough to replicate, and the empirical claims can be tested directly once the full results and ablations are examined.

Referee Report

3 major / 2 minor

Summary. The paper proposes LEMON, an LLM-based system for generating executable multi-agent orchestration specifications that include roles, duties, capacities, and dependency structures. It trains the orchestrator by augmenting the GRPO objective with a localized counterfactual reinforcement learning signal, where individual orchestration fields are edited to compute reward contrasts applied only to the edited spans. The method is evaluated on six benchmarks (MMLU, GSM8K, AQuA, MultiArith, SVAMP, HumanEval), claiming state-of-the-art performance among multi-agent orchestration methods.

Significance. If the counterfactual approach successfully provides localized credit assignment superior to standard execution-level feedback, this work could significantly advance the automated optimization of multi-agent LLM systems by enabling more precise and efficient orchestration design, with potential applications in complex reasoning and coding tasks.

major comments (3)

The assumption that single-field edits (role, capacity, or dependency) produce reward contrasts that can be cleanly attributed to the edited span is load-bearing but potentially violated by cascading effects in agent dependencies and execution paths; no analysis of how frequently such edits alter downstream sequences is provided.
The central empirical claim of SOTA performance lacks supporting details on baselines, statistical tests, ablation studies (e.g., localized vs. global reward), or error bars, preventing verification of the improvement from the counterfactual signal.
Table or results section reporting performance on the six benchmarks should include comparisons with error bars and significance tests to substantiate the SOTA claim.

minor comments (2)

The abstract mentions 'our code is available' but the link is anonymous; ensure the full paper provides a permanent link post-review.
Ensure all acronyms like GRPO are defined on first use in the main text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on the counterfactual reinforcement learning approach and the empirical evaluation. We address each major comment below and will incorporate revisions to strengthen the manuscript's clarity and rigor.

read point-by-point responses

Referee: The assumption that single-field edits (role, capacity, or dependency) produce reward contrasts that can be cleanly attributed to the edited span is load-bearing but potentially violated by cascading effects in agent dependencies and execution paths; no analysis of how frequently such edits alter downstream sequences is provided.

Authors: We agree that cascading effects represent a potential limitation of the localized counterfactual signal. While the design isolates edits to specific spans and applies contrasts only to those tokens, we acknowledge that downstream execution paths may change in some cases. In the revised manuscript, we will add a dedicated analysis section quantifying the frequency of downstream sequence alterations following single-field edits (role, capacity, and dependency) across the six benchmarks, including statistics on how often such edits propagate beyond the edited span. revision: yes
Referee: The central empirical claim of SOTA performance lacks supporting details on baselines, statistical tests, ablation studies (e.g., localized vs. global reward), or error bars, preventing verification of the improvement from the counterfactual signal.

Authors: We will revise the experimental section to provide full details on all baseline implementations, including exact prompting and training configurations for comparison methods. We will add ablation studies directly comparing the localized counterfactual GRPO objective against a global execution-level reward variant. Results will be reported with error bars (standard deviation over multiple random seeds) and statistical significance tests (paired t-tests with p-values) to substantiate the gains attributable to the counterfactual signal. revision: yes
Referee: Table or results section reporting performance on the six benchmarks should include comparisons with error bars and significance tests to substantiate the SOTA claim.

Authors: The results table and accompanying text will be updated in the revision to include error bars computed over five independent runs for all methods and benchmarks. We will also report p-values from paired statistical tests for key comparisons against the strongest baselines, ensuring the SOTA claims are supported by verifiable evidence. revision: yes

Circularity Check

0 steps flagged

No circularity: LEMON applies existing RL ideas to orchestration without self-referential derivations

full rationale

The paper frames LEMON as an augmentation of the standard GRPO objective with counterfactual edits on orchestration fields (role, capacity, dependency). No equations, derivations, or self-citations are presented that reduce the claimed performance gains or credit-assignment mechanism to a fitted parameter defined by the same data or to a prior result whose validity depends on the current work. The method is described as an application of counterfactual RL to a new specification format, with experiments on external benchmarks. This is self-contained against external benchmarks and contains no load-bearing steps that collapse by construction. Score 0 is the appropriate finding per the guidelines for papers whose central claim does not reduce to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the approach rests on standard reinforcement-learning assumptions and the capability of LLMs to follow orchestration instructions; no explicit free parameters, new entities, or ad-hoc axioms are named.

axioms (1)

domain assumption Standard assumptions of policy-gradient reinforcement learning hold for the GRPO objective when applied to orchestration decisions.
The training procedure is described as an augmentation of GRPO without stating deviations from its usual convergence or credit-assignment properties.

pith-pipeline@v0.9.0 · 5545 in / 1225 out tokens · 39629 ms · 2026-05-15T02:10:02.859340+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 8 internal anchors

[1]

Autogen: Enabling next-gen llm applications via multi-agent conversations

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: Enabling next-gen llm applications via multi-agent conversations. InFirst conference on language modeling, 2024

work page 2024
[2]

Camel: Communicative agents for" mind" exploration of large language model society.Advances in neural information processing systems, 36:51991–52008, 2023

Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society.Advances in neural information processing systems, 36:51991–52008, 2023

work page 2023
[3]

Metagpt: Meta programming for a multi-agent collaborative framework

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al. Metagpt: Meta programming for a multi-agent collaborative framework. InThe twelfth international conference on learning representations, 2023

work page 2023
[4]

Large Language Model based Multi-Agents: A Survey of Progress and Challenges

Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[5]

Improv- ing factuality and reasoning in language models through multiagent debate

Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improv- ing factuality and reasoning in language models through multiagent debate. InForty-first international conference on machine learning, 2024

work page 2024
[6]

Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023

work page 2023
[7]

Gptswarm: Language agents as optimizable graphs

Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. Gptswarm: Language agents as optimizable graphs. InForty-first International Conference on Machine Learning, 2024

work page 2024
[8]

G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

Guibin Zhang, Yanwei Yue, Xiangguo Sun, Guancheng Wan, Miao Yu, Junfeng Fang, Kun Wang, Tianlong Chen, and Dawei Cheng. G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

work page arXiv 2024
[9]

Cut the crap: An economical communication pipeline for llm-based multi-agent systems.arXiv preprint arXiv:2410.02506, 2024

Guibin Zhang, Yanwei Yue, Zhixun Li, Sukwon Yun, Guancheng Wan, Kun Wang, Dawei Cheng, Jeffrey Xu Yu, and Tianlong Chen. Cut the crap: An economical communication pipeline for llm-based multi-agent systems.arXiv preprint arXiv:2410.02506, 2024

work page arXiv 2024
[10]

AFlow: Automating Agentic Workflow Generation

Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow generation.arXiv preprint arXiv:2410.10762, 2024

work page internal anchor Pith review arXiv 2024
[11]

Learning to Orchestrate Agents in Natural Language with the Conductor

Stefan Nielsen, Edoardo Cetin, Peter Schwendeman, Qi Sun, Jinglue Xu, and Yujin Tang. Learning to orchestrate agents in natural language with the conductor.arXiv preprint arXiv:2512.04388, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[12]

Workflowllm: Enhancing workflow orchestration capability of large language models.arXiv preprint arXiv:2411.05451, 2024

Shengda Fan, Xin Cong, Yuepeng Fu, Zhong Zhang, Shuyan Zhang, Yuanwei Liu, Yesai Wu, Yankai Lin, Zhiyuan Liu, and Maosong Sun. Workflowllm: Enhancing workflow orchestration capability of large language models.arXiv preprint arXiv:2411.05451, 2024

work page arXiv 2024
[13]

Multi-agent architecture search via agentic supernet.arXiv preprint arXiv:2502.04180, 2025

Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, Lei Bai, and Xiang Wang. Multi-agent architecture search via agentic supernet.arXiv preprint arXiv:2502.04180, 2025

work page arXiv 2025
[14]

Agentbalance: Backbone-then-topology design for cost-effective multi-agent systems under budget constraints.arXiv preprint arXiv:2512.11426, 2025

Shuowei Cai, Yansong Ning, and Hao Liu. Agentbalance: Backbone-then-topology design for cost-effective multi-agent systems under budget constraints.arXiv preprint arXiv:2512.11426, 2025

work page arXiv 2025
[15]

Assemble your crew: Automatic multi-agent communication topology design via autoregressive graph generation

Shiyuan Li, Yixin Liu, Qingsong Wen, Chengqi Zhang, and Shirui Pan. Assemble your crew: Automatic multi-agent communication topology design via autoregressive graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 23142–23150, 2026. 10

work page 2026
[16]

Ofa-mas: One-for-all multi-agent system topology design based on mixture-of-experts graph generative models

Shiyuan Li, Yixin Liu, Yu Zheng, Mei Li, Quoc Viet Hung Nguyen, and Shirui Pan. Ofa-mas: One-for-all multi-agent system topology design based on mixture-of-experts graph generative models. InProceedings of the ACM Web Conference 2026, pages 1333–1344, 2026

work page 2026
[17]

Automated Design of Agentic Systems

Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems.arXiv preprint arXiv:2408.08435, 2024

work page internal anchor Pith review arXiv 2024
[18]

Difficulty-aware agentic orchestration for query-specific multi-agent workflows

Jinwei Su, Qizhen Lan, Yinghui Xia, Lifan Sun, Weiyou Tian, Tianyu Shi, and Lewei He. Difficulty-aware agentic orchestration for query-specific multi-agent workflows. InProceedings of the ACM Web Conference 2026, pages 2060–2070, 2026

work page 2026
[19]

Masrouter: Learning to route llms for multi-agent systems

Yanwei Yue, Guibin Zhang, Boyang Liu, Guancheng Wan, Kun Wang, Dawei Cheng, and Yiyan Qi. Masrouter: Learning to route llms for multi-agent systems. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15549–15572, 2025

work page 2025
[20]

arXiv preprint arXiv:2602.20229 , year =

Tianjun Yao, Zhaoyi Li, and Zhiqiang Shen. Hieramas: Optimizing intra-node llm mixtures and inter-node topology for multi-agent systems.arXiv preprint arXiv:2602.20229, 2026

work page arXiv 2026
[21]

Toolorchestra: Elevating intelligence via efficient model and tool orchestration.arXiv preprint arXiv:2511.21689, 2025

Hongjin Su, Shizhe Diao, Ximing Lu, Mingjie Liu, Jiacheng Xu, Xin Dong, Yonggan Fu, Peter Belcak, Hanrong Ye, Hongxu Yin, et al. Toolorchestra: Elevating intelligence via efficient model and tool orchestration.arXiv preprint arXiv:2511.21689, 2025

work page arXiv 2025
[22]

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

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, 2025

work page arXiv 2025
[23]

Topoweaver-r1: Reinforcing difficulty-aware topology evolution in multi-agent competition-level code generation

Siyu Wang, Ruotian Lu, Zhihao Yang, Yuchao Wang, Lei Xu, Qimin Xu, Guojun Yin, Cailian Chen, Xinping Guan, et al. Topoweaver-r1: Reinforcing difficulty-aware topology evolution in multi-agent competition-level code generation

work page
[24]

Fd-magrpo: Functionality-driven multi-agent group relative policy optimization for analog-ldo sizing

Haoning Jiang, Han Wu, Zhuoli Ouyang, Ziheng Wang, Tinghuan Chen, and Junmin Jiang. Fd-magrpo: Functionality-driven multi-agent group relative policy optimization for analog-ldo sizing. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 22310–22317, 2026

work page 2026
[25]

Graph-grpo: Training graph flow models with reinforcement learning.arXiv preprint arXiv:2603.10395, 2026

Baoheng Zhu, Deyu Bo, Delvin Ce Zhang, and Xiao Wang. Graph-grpo: Training graph flow models with reinforcement learning.arXiv preprint arXiv:2603.10395, 2026

work page arXiv 2026
[26]

Training Verifiers to Solve Math Word Problems

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[27]

Solving general arithmetic word problems

Subhro Roy and Dan Roth. Solving general arithmetic word problems. InProceedings of the 2015 conference on empirical methods in natural language processing, pages 1743–1752, 2015

work page 2015
[28]

Arkil Patel, Satwik Bhattamishra, and Navin Goyal. Are nlp models really able to solve simple math word problems? InProceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 2080–2094, 2021

work page 2021
[29]

Program induction by rationale generation: Learning to solve and explain algebraic word problems

Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. InProceedings of the 55th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 158–167, 2017

work page 2017
[30]

Measuring Massive Multitask Language Understanding

Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2009
[31]

Evaluating Large Language Models Trained on Code

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021. 11

work page internal anchor Pith review Pith/arXiv arXiv 2021
[32]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

work page 2022
[33]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[34]

Limitations

Zhexuan Wang, Yutong Wang, Xuebo Liu, Liang Ding, Miao Zhang, Jie Liu, and Min Zhang. Agentdropout: Dynamic agent elimination for token-efficient and high-performance llm-based multi-agent collaboration. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 24013–24035, 2025. 12 A Example ...

work page 2025
[35]

Therefore, IRB approval or equivalent ethics review is not applicable

Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

work page