Pith. sign in

REVIEW 4 major objections 6 minor 40 references

By encoding both the evolving problem state and each agent's expertise, STRMAC picks the agent best matched to the current step, reporting accuracy gains up to 23.8% while cutting token use and training-path search overhead by up to 90.1%.

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

2026-08-04 00:12 UTC pith:HDL3P5FC

load-bearing objection Interesting routing idea with big token-efficiency claims, but the core equation is unimplementable as written and the empirical reporting is too thin to verify. the 4 major comments →

arxiv 2511.02200 v2 pith:HDL3P5FC submitted 2025-11-04 cs.AI

Optimal-Agent-Selection: State-Aware Routing Framework for Efficient Multi-Agent Collaboration

Jingbo Wang , Sendong Zhao , Haochun Wang , Yuzheng Fan , Ting Liu This is my paper
classification cs.AI
keywords multi-agent systemsLLM routingstate-aware routingagent selectioncontrastive learningself-evolving data generationtoken efficiencycollaborative reasoning
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 argues that in multi-agent LLM collaboration, the order in which agents act matters as much as which agents exist, and that the best agent for the next step can be predicted from the evolving state. It introduces STRMAC, which encodes the current problem state and each agent's expertise as vectors and selects the agent with the highest cosine similarity; a contrastive objective trains the state encoder to prefer the agent that actually produced the best correct outcome. To make training tractable, it mines execution paths with early pruning and router-guided search, reducing data-collection overhead by up to 90.1% compared to exhaustive enumeration. If correct, the framework would replace rigid pipelines and static collaboration graphs with a cheap, per-step routing decision, improving both accuracy and token efficiency.

Core claim

STRMAC claims that multi-agent collaboration should be treated as a routing problem rather than a fixed pipeline: at each step, a trained encoder embeds the current state (the question plus interaction history), while each agent's private knowledge is pre-embedded by an LLM; the agent whose embedding has the highest cosine similarity to the state is selected to act. A contrastive loss aligns state representations with the agent that produced the best and cheapest correct outcome. The paper further claims that training paths can be mined efficiently by early pruning once a correct answer is found and by router-guided exploration, cutting data-collection overhead by up to 90.1% relative to exh

What carries the argument

The central mechanism is the state–agent compatibility score: the router encoder produces a latent state z_t; each agent has a fixed embedding e_i from its LLM; compatibility is cosine similarity between z_t and e_i; a softmax over agents yields probabilities, and the highest-score agent is chosen to act next. This one-step greedy selection is trained with a contrastive loss that pulls the state embedding toward the optimal agent's embedding and pushes it away from the others. The companion mechanism is self-evolving data generation: search paths are pruned as soon as a correct answer is reached, alternative valid paths are kept as auxiliary training signals, and after an initial router is t

Load-bearing premise

The whole routing score rests on the assumption that the state encoder's embedding and each agent's embedding live in a shared vector space where cosine similarity is meaningful; the paper does not describe how embeddings of different dimensions are aligned, so as written that compatibility score is undefined.

What would settle it

Take a trained STRMAC system and compute the routing score exactly as defined: if the state and agent embeddings have different dimensions, the cosine similarity cannot be evaluated; after fixing a projection, compare routing accuracy against a random-agent baseline on the same benchmarks—the central claim stands only if the projection-controlled system still beats fixed pipelines by the reported margins.

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

If this is right

  • Fixed multi-agent pipelines and static task graphs are unnecessary; per-step state matching can outperform them in both accuracy and cost.
  • Token efficiency becomes a trainable objective: routing on compatibility can cut inference cost to roughly 5–20% of existing approaches while keeping or improving accuracy.
  • A router trained on data from a smaller model transfers to larger models, lowering the cost of collecting training trajectories.
  • The self-evolving search makes multi-agent system training feasible when the number of agents makes exhaustive path enumeration impossible.
  • The learned routing policy is interpretable: it tends to select the agent with the most comprehensive information first, then consults supplementary agents only when needed.

Where Pith is reading between the lines

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

  • The one-step greedy compatibility rule may be myopic; a natural extension is a learned scorer or a reinforcement-learning objective that anticipates the value of future steps, which could help when short-term similarity is misleading.
  • The same embedding-compatibility idea could route within a single model—across layers, experts, or reasoning strategies—rather than only across distinct agents.
  • If the reported dimension mismatch between the state encoder and agent embeddings is resolved with a projection, the method could be adapted to heterogeneous agent pools with different underlying encoders, widening its practical reach.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes STRMAC, a state-aware routing framework for multi-agent LLM collaboration. At each interaction step, a lightweight trainable encoder (mDeBERTaV3-base) embeds the accumulated state, and the router selects the agent whose fixed embedding (derived from the base LLM's embedding layer) has the highest cosine similarity with the state embedding. The router is trained with a contrastive loss to align states with the agent that produced the optimal path, where optimality is defined as correctness plus minimal token consumption. To avoid exhaustive enumeration of agent permutations, the authors introduce a self-evolving data-generation pipeline combining solution-aware pruning, router-guided top-k expansion, and iterative retraining. Experiments on two self-constructed benchmarks, PDDP and EBFC, report state-of-the-art accuracy and substantially lower token usage relative to six baselines, plus additional transfer experiments on GPT-4o.

Significance. The high-level idea — dynamically routing to a single most suitable agent at each step, using an explicit state representation and fixed agent expertise embeddings — is timely and, if implemented correctly, would be a practical contribution to multi-agent efficiency. The contrastive objective and the self-evolving data-collection scheme are conceptually appealing, and the reported token savings are large. The paper also provides a reasonable breadth of baselines and a cost-adjusted metric. However, the central routing equation is underspecified to the point of being non-implementable: the state and agent embeddings are drawn from models with incompatible dimensions, and no projection or pooling is described. Since every reported accuracy and token number flows through this equation, the experimental results cannot be verified from the manuscript as written. The absence of dataset sizes, train/test splits, error bars, and significance testing further weakens the empirical claims.

major comments (4)
  1. [§3.2, Eq. (3.2); §4.1; §5.3] The core compatibility score is dimensionally undefined. The state encoder is mDeBERTaV3-base (86M parameters, hidden dimension 768), while agent embeddings are stated to come from the embedding layers of Llama-3.1-70B (4096), Llama-3.1-8B (4096), and Qwen2.5-32B (5120). Cosine similarity between vectors in R^768 and R^4096 (or R^5120) is not defined. No projection, linear layer, pooling, or dimension-matching mechanism is described in §3.2 or §3.3. Moreover, §3.3 explicitly states that only router-encoder parameters are updated and agent embeddings remain fixed, which rules out a silently learned projection. Section 5.3 compounds the problem by substituting Llama-3.1-70B embeddings for GPT-4o, again producing 8192-dimensional vectors. As written, the method cannot be re-implemented, and the results in Tables 1 and 4 are unverifiable. The authors must specify the exact projection/pooling
  2. [§4.1, Tables 1, 3, 4] No dataset sizes, train/validation/test split sizes, number of evaluation instances, or number of independent runs are reported. None of the accuracy comparisons are accompanied by error bars, confidence intervals, or significance tests. Many headline differences are small in absolute terms (e.g., PDDP Llama-70B: STRMAC 64.0 vs. IG-MAS 62.4; GPT-4o PDDP: 64.4 vs. IG-MAS 63.4). Given that the benchmarks are constructed by the authors according to their own protocol, the absence of variance information makes it impossible to assess whether the claimed improvements are robust. Please report the evaluation set size, repeated-run variance, and appropriate statistical tests for the main comparisons.
  3. [§5.1, Table 2; Abstract] The claim of 'reducing data collection overhead by up to 90.1%' is based on the number of sampled execution paths relative to exhaustive search. However, the actual overhead of the proposed pipeline also includes the cost of training and iteratively retraining the router, the cost of scoring candidate paths by executing the full multi-agent system, and the cost of router-guided expansion. Table 2 reports only path counts, not wall-clock time, GPU-hours, or token consumption during data generation. The efficiency claim as stated is therefore not substantiated. Please report end-to-end data-collection cost or clarify that the 90.1% figure refers only to path-sample counts.
  4. [§4.1, §3.2] Even after resolving the dimension mismatch, the construction of agent embeddings is underspecified. The paper says agent embeddings are obtained from 'the embedding layers' of the base LLMs, but for decoder-only models the embedding layer maps token IDs to token embeddings. It is unclear whether the agent embedding is a single-token vector, a mean/max pooling of the prompt tokens, the last-token hidden state, or something else. Since the agent embedding is the fixed target of the contrastive loss, this choice materially affects the router's behavior. Please specify the exact extraction procedure, including the prompt template used for each agent.
minor comments (6)
  1. [Eq. (3.3)] The sentence after the equation is missing a period: 'where⟨·,·⟩represents cosine similarity' should end with a period.
  2. [Table 1, EBFC, MDAgents, Llama3.1-8B row] The token entry '0.872.327817.9' appears to be a formatting typo; it should likely be '8723.2' or similar. Please correct.
  3. [§4.1, Eq. (4.1)] CAS depends on two free parameters, mu and c, set to 0.1 and 1000 without sensitivity analysis. Since CAS is used to rank methods, a short sensitivity study would clarify whether the qualitative conclusions are robust to reasonable perturbations of these values.
  4. [§1, Figure 1] Figure 1's caption refers to 'five agents (A-E)' but does not state which task or dataset the accuracy comparison is drawn from. Adding this context would improve interpretability.
  5. [§3.4, Eq. (3.5)] The path-scoring equation assigns −∞ to all incorrect paths. This is a reasonable simplification, but the paper does not discuss ties among correct paths with the same token count, which could occur in practice. Please explain how ties are broken.
  6. [§5.2, Table 3] The generalization claim is based on comparing accuracy differences of about 1–2 points. Given the absence of error bars (noted above), the claim that routers trained on Llama-8B data are 'comparable' to those trained on same-model data should be tempered.

Circularity Check

0 steps flagged

No circular reduction found: the router is trained on ground-truth-labeled paths and evaluated on held-out MIMIC-III/AMBIFC instances; the main risks (self-built benchmarks, hand-set CAS constants, vector-dimension mismatch) are validity concerns, not equations reducing to their inputs.

full rationale

The derivation chain is not circular. The contrastive training objective (Eq. 3.3) uses agent embeddings fixed from LLM layers and state embeddings from a trainable encoder; the supervision (Eq. 3.5) is defined by ground-truth answer correctness and measured token count, not by a fitted parameter that is later re-presented as a prediction. The reported accuracy numbers come from actually running the selected agents on held-out instances of PDDP (MIMIC-III) and EBFC (AMBIFC), so the central claim is externally checkable. Using (Wang et al. 2025) for the benchmark protocols and the IG-MAS baseline is a same-group citation, and the CAS metric (Eq. 4.1) is hand-parameterized (mu=0.1, c=1000); both affect the strength and interpretability of the evaluation, but neither makes a prediction equivalent to its input. The most serious technical defect in the paper—Eq. 3.2 compares mDeBERTaV3-base state embeddings (768-d) with Llama/Qwen embedding-layer vectors (4096-8192-d) without a projection—is an implementation/correctness issue, not a circularity: a cosine similarity that is undefined is not a result that reduces to its own assumptions. Therefore no circular step meets the quote-and-reduction bar.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on a handful of free parameters (CAS scoring, data-mixture split, unstated top-k) and domain assumptions about agent solvability, the optimality criterion, and the sufficiency of frozen embeddings. No new physical entities are introduced. The most fragile assumption is the unstated dimension-matching between state and agent embeddings.

free parameters (5)
  • CAS cost sensitivity mu = 0.1
    Eq 4.1; hand-set to penalize token usage; directly controls the CAS rankings in Table 1.
  • CAS normalization constant c = 1000
    Eq 4.1; hand-set so that every additional 1000 tokens reduces the score by a factor of exp(-0.1).
  • Router-guided expansion top-k = not specified
    Section 3.4; the number of agents the router proposes for expansion is never given, yet it controls data-generation cost and path diversity.
  • Initial pruning/router-guided data mixture = 20% / 80%
    Section 5.1; the split between solution-aware pruning and router-guided exploration is chosen by hand.
  • Learning rate = 1e-4 to 5e-5
    Section 4.1; tuned on validation per configuration; standard hyperparameter but not reported per model.
axioms (5)
  • domain assumption Each agent A_i can solve the complete task on its own; the router merely chooses who acts next.
    Section 3.1 states 'each have the ability to solve the complete task on their own.' If no single agent can contribute correct reasoning, greedy per-step selection would fail; PDDP ablation (Table 5) shows only AgentBHC is a strong solo performer.
  • ad hoc to paper Optimal path is defined as any correct path with minimal token count; all incorrect paths are equally bad (-inf).
    Eq 3.5 fixes the training labels for the contrastive router; it is a convenience criterion, not derived from first principles, and it conflates efficiency with optimality.
  • domain assumption Frozen LLM embeddings e_i are sufficient to characterize agent suitability at every step.
    Section 3.2 fixes agent embeddings from a separate LLM and only trains the state encoder; the paper provides no ablation on learning agent embeddings jointly.
  • domain assumption A partially trained router's top-k suggestions are a reliable guide for exploring the execution tree.
    Section 3.4's router-guided expansion assumes early router probabilities point to promising branches; if the early router is poor, pruning can discard good paths.
  • standard math Contrastive InfoNCE loss with softmax over cosine similarities yields a calibrated probability over agents.
    Standard ML background; not proven in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 11851 in / 14754 out tokens · 150354 ms · 2026-08-04T00:12:32.989939+00:00 · methodology

0 comments
read the original abstract

The emergence of multi-agent systems powered by large language models (LLMs) has unlocked new frontiers in complex task-solving, enabling diverse agents to integrate unique expertise, collaborate flexibly, and address challenges unattainable for individual models. However, the full potential of such systems is hindered by rigid agent scheduling and inefficient coordination strategies that fail to adapt to evolving task requirements. In this paper, we propose STRMAC, a state-aware routing framework designed for efficient collaboration in multi-agent systems. Our method separately encodes interaction history and agent knowledge to power the router, which adaptively selects the most suitable single agent at each step for efficient and effective collaboration. Furthermore, we introduce a self-evolving data generation approach that accelerates the collection of high-quality execution paths for efficient system training. Experiments on challenging collaborative reasoning benchmarks demonstrate that our method achieves state-of-the-art performance, achieving up to 23.8% improvement over baselines and reducing data collection overhead by up to 90.1% compared to exhaustive search.

Figures

Figures reproduced from arXiv: 2511.02200 by Haochun Wang, Jingbo Wang, Sendong Zhao, Ting Liu, Yuzheng Fan.

Figure 1
Figure 1. Figure 1: Accuracy comparison of two work paths for five [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of STRMAC framework. Each agent’s private context [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distribution (bar, left axis) and accuracy (line, [PITH_FULL_IMAGE:figures/full_fig_p007_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

40 extracted references · 25 linked inside Pith

  1. [1]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  2. [2]

    Chen, L.; Zaharia, M.; and Zou, J. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176

    Pith/arXiv arXiv 2023

  3. [3]

    Chen, S.; Jiang, W.; Lin, B.; Kwok, J.; and Zhang, Y. 2024. Routerdc: Query-based router by dual contrastive learning for assembling large language models. Advances in Neural Information Processing Systems, 37: 66305--66328

    2024

  4. [4]

    Chen, W.; Su, Y.; Zuo, J.; Yang, C.; Yuan, C.; Qian, C.; Chan, C.-M.; Qin, Y.; Lu, Y.; Xie, R.; et al. 2023. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2(4): 6

    Pith/arXiv arXiv 2023

  5. [5]

    Dang, Y.; Qian, C.; Luo, X.; Fan, J.; Xie, Z.; Shi, R.; Chen, W.; Yang, C.; Che, X.; Tian, Y.; et al. 2025. Multi-Agent Collaboration via Evolving Orchestration. arXiv preprint arXiv:2505.19591

    arXiv 2025

  6. [6]

    Deng, X.; Gu, Y.; Zheng, B.; Chen, S.; Stevens, S.; Wang, B.; Sun, H.; and Su, Y. 2023. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36: 28091--28114

    2023

  7. [7]

    V.; and Awadallah, A

    Ding, D.; Mallick, A.; Wang, C.; Sim, R.; Mukherjee, S.; Ruhle, V.; Lakshmanan, L. V.; and Awadallah, A. H. 2024. Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618

    Pith/arXiv arXiv 2024

  8. [8]

    B.; and Mordatch, I

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

    2023

  9. [9]

    Glockner, M.; Stali \=u nait \.e , I.; Thorne, J.; Vallejo, G.; Vlachos, A.; and Gurevych, I. 2024. Ambifc: Fact-checking ambiguous claims with evidence. Transactions of the Association for Computational Linguistics, 12: 1--18

    2024

  10. [10]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    Pith/arXiv arXiv 2024

  11. [11]

    He, P.; Liu, X.; Gao, J.; and Chen, W. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654

    Pith/arXiv arXiv 2020

  12. [12]

    Hong, S.; Lin, Y.; Liu, B.; Liu, B.; Wu, B.; Zhang, C.; Wei, C.; Li, D.; Chen, J.; Zhang, J.; et al. 2024. Data interpreter: An llm agent for data science. arXiv preprint arXiv:2402.18679

    Pith/arXiv arXiv 2024

  13. [13]

    M.; Luck, M.; Bu, Q.; Qing, Y.; and Cui, H

    Huang, D.; Zhang, J. M.; Luck, M.; Bu, Q.; Qing, Y.; and Cui, H. 2023. Agentcoder: Multi-agent-based code generation with iterative testing and optimisation. arXiv preprint arXiv:2312.13010

    Pith/arXiv arXiv 2023

  14. [14]

    Hui, B.; Yang, J.; Cui, Z.; Yang, J.; Liu, D.; Zhang, L.; Liu, T.; Zhang, J.; Yu, B.; Lu, K.; et al. 2024. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186

    Pith/arXiv arXiv 2024

  15. [15]

    E.; Pollard, T

    Johnson, A. E.; Pollard, T. J.; Shen, L.; Lehman, L.-w. H.; Feng, M.; Ghassemi, M.; Moody, B.; Szolovits, P.; Anthony Celi, L.; and Mark, R. G. 2016. MIMIC-III, a freely accessible critical care database. Scientific data, 3(1): 1--9

    2016

  16. [16]

    B.; Becker, J.; Wahle, J

    Kaesberg, L. B.; Becker, J.; Wahle, J. P.; Ruas, T.; and Gipp, B. 2025. Voting or consensus? decision-making in multi-agent debate. arXiv preprint arXiv:2502.19130

    arXiv 2025

  17. [17]

    S.; Xu, X.; McDuff, D.; Lee, H.; Ghassemi, M.; Breazeal, C.; and Park, H

    Kim, Y.; Park, C.; Jeong, H.; Chan, Y. S.; Xu, X.; McDuff, D.; Lee, H.; Ghassemi, M.; Breazeal, C.; and Park, H. W. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems, 37: 79410--79452

    2024

  18. [18]

    Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024 a . Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    Pith/arXiv arXiv 2024

  19. [19]

    Liu, Z.; Zhang, Y.; Li, P.; Liu, Y.; and Yang, D. 2024 b . A dynamic LLM-powered agent network for task-oriented agent collaboration. In First Conference on Language Modeling

    2024

  20. [20]

    Lu, K.; Yuan, H.; Lin, R.; Lin, J.; Yuan, Z.; Zhou, C.; and Zhou, J. 2023. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692

    Pith/arXiv arXiv 2023

  21. [21]

    E.; Kadous, M

    Ong, I.; Almahairi, A.; Wu, V.; Chiang, W.-L.; Wu, T.; Gonzalez, J. E.; Kadous, M. W.; and Stoica, I. 2024. RouteLLM: Learning to Route LLMs from Preference Data. In The Thirteenth International Conference on Learning Representations

    2024

  22. [22]

    OpenAI. 2024. GPT-4o System Card. Computing Research Repository, abs/2410.21276

    Pith/arXiv arXiv 2024

  23. [23]

    Y.; Ma, X.; and Kim, J

    Park, J.; Min, B.; Son, K.; Song, J. Y.; Ma, X.; and Kim, J. 2023. Choicemates: Supporting unfamiliar online decision-making with multi-agent conversational interactions. arXiv preprint arXiv:2310.01331

    arXiv 2023

  24. [24]

    Pratt, J. W. 1978. Risk aversion in the small and in the large. In Uncertainty in economics, 59--79. Elsevier

    1978

  25. [25]

    Qian, C.; Liu, W.; Liu, H.; Chen, N.; Dang, Y.; Li, J.; Yang, C.; Chen, W.; Su, Y.; Cong, X.; et al. 2023. Chatdev: Communicative agents for software development. arXiv preprint arXiv:2307.07924

    Pith/arXiv arXiv 2023

  26. [26]

    Qian, C.; Xie, Z.; Wang, Y.; Liu, W.; Dang, Y.; Du, Z.; Chen, W.; Yang, C.; Liu, Z.; and Sun, M. 2024. Scaling large-language-model-based multi-agent collaboration. arXiv preprint arXiv:2406.07155

    Pith/arXiv arXiv 2024

  27. [27]

    E.; Lv, C.; and Chen, H

    Qiao, S.; Zhang, N.; Fang, R.; Luo, Y.; Zhou, W.; Jiang, Y. E.; Lv, C.; and Chen, H. 2024. AutoAct: Automatic agent learning from scratch for QA via self-planning. arXiv preprint arXiv:2401.05268

    Pith/arXiv arXiv 2024

  28. [28]

    Tang, X.; Zou, A.; Zhang, Z.; Li, Z.; Zhao, Y.; Zhang, X.; Cohan, A.; and Gerstein, M. 2023. Medagents: Large language models as collaborators for zero-shot medical reasoning. arXiv preprint arXiv:2311.10537

    Pith/arXiv arXiv 2023

  29. [29]

    Touvron, H.; Lavril, T.; Izacard, G.; Martinet, X.; Lachaux, M.-A.; Lacroix, T.; Rozi \`e re, B.; Goyal, N.; Hambro, E.; Azhar, F.; et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    Pith/arXiv arXiv 2023

  30. [30]

    Wang, G.; Xie, Y.; Jiang, Y.; Mandlekar, A.; Xiao, C.; Zhu, Y.; Fan, L.; and Anandkumar, A. 2023. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291

    Pith/arXiv arXiv 2023

  31. [31]

    Wang, H.; Zhao, S.; Wang, J.; Qiang, Z.; Qin, B.; and Liu, T. 2025. Beyond Frameworks: Unpacking Collaboration Strategies in Multi-Agent Systems. arXiv preprint arXiv:2505.12467

    Pith/arXiv arXiv 2025

  32. [32]

    Wang, J.; Wang, J.; Athiwaratkun, B.; Zhang, C.; and Zou, J. 2024 a . Mixture-of-agents enhances large language model capabilities. arXiv preprint arXiv:2406.04692

    Pith/arXiv arXiv 2024

  33. [33]

    Wang, Q.; Wang, Z.; Su, Y.; Tong, H.; and Song, Y. 2024 b . Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key? arXiv preprint arXiv:2402.18272

    Pith/arXiv arXiv 2024

  34. [34]

    X.; and Chen, T

    Zhang, G.; Yue, Y.; Li, Z.; Yun, S.; Wan, G.; Wang, K.; Cheng, D.; Yu, J. X.; and Chen, T. 2024 a . Cut the crap: An economical communication pipeline for llm-based multi-agent systems. arXiv preprint arXiv:2410.02506

    Pith/arXiv arXiv 2024

  35. [35]

    Zhang, G.; Yue, Y.; Sun, X.; Wan, G.; Yu, M.; Fang, J.; Wang, K.; Chen, T.; and Cheng, D. 2024 b . G-designer: Architecting multi-agent communication topologies via graph neural networks. arXiv preprint arXiv:2410.11782

    Pith/arXiv arXiv 2024

  36. [36]

    Zhang, T.; Li, D.; Chen, Q.; Wang, C.; and He, X. 2025. BELLE: A Bi-Level Multi-Agent Reasoning Framework for Multi-Hop Question Answering. arXiv preprint arXiv:2505.11811

    Pith/arXiv arXiv 2025

  37. [37]

    Zhang, Y.-K.; Zhan, D.-C.; and Ye, H.-J. 2025. Capability instruction tuning: A new paradigm for dynamic llm routing. arXiv preprint arXiv:2502.17282

    Pith/arXiv arXiv 2025

  38. [38]

    Zhuge, M.; Wang, W.; Kirsch, L.; Faccio, F.; Khizbullin, D.; and Schmidhuber, J. 2024. Gptswarm: Language agents as optimizable graphs. In Forty-first International Conference on Machine Learning

    2024

  39. [39]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  40. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...