Pith. sign in

REVIEW 3 major objections 5 minor 43 references

OSC: Cognitive Orchestration through Dynamic Knowledge Alignment in Multi-Agent LLM Collaboration

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A learned collaboration layer that lets each LLM agent model what its teammates know lifts AlpacaEval 2.0 win rate to 81.4% and cuts communication overhead.

desk verdict The OSC framework is a real architectural idea, but the headline AlpacaEval result is in-sample and the promised reasoning benchmarks are missing, so the central claim is unsupported as reported. read the letter →

arxiv 2509.04876 v1 pith:3K4KJDLW submitted 2025-09-05 cs.AI

Jusheng Zhang , Yijia Fan , Kaitong Cai , Xiaofei Sun , Keze Wang This is my paper
classification cs.AI
keywords multi-agentLLMsystemscollaboratorknowledgemodelcognitivegapanalysisadaptivecommunicationpolicyreinforcementlearningPPOAlpacaEval2.0agentcollaboration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the main bottleneck in multi-agent LLM systems is neither which experts are chosen nor how their outputs are merged, but how they talk to one another. It introduces OSC, an intermediate collaboration layer in which each agent maintains a learned Collaborator Knowledge Model (CKM) of every teammate's current knowledge, confidence, and task understanding. A learned cognitive-gap function compares that model with the agent's own state, and a PPO-trained communication policy decides whom to address, which gap to close, what objective to pursue, and in what style. The headline result is an 81.4% length-controlled win rate on AlpacaEval 2.0, above KABB (77.9%) and MoA (68.1%), with fewer rounds and tokens than comparison systems. A sympathetic reader would care because it reframes multi-agent performance as a communication-design problem and offers a trainable mechanism for making agents' messages contingent on collaborators' inferred mental states.

What carries the argument

The Collaborator Knowledge Model (CKM) is the load-bearing mechanism: a per-agent-pair latent state, produced by a Transformer encoder and updated by a GRU, that encodes agent i's evolving belief about agent j's knowledge, confidence, and task understanding. It carries the argument because every downstream decision—gap analysis, target selection, communication objective, and style choice—is a function of these learned beliefs, and because the CKM, the gap function, and the policy are fine-tuned end-to-end from the same task reward, the representations are shaped by whether they lead to successful collaboration.

What would settle it

Take the 805 AlpacaEval instructions, remove the 160-instruction development subset from all PPO training and hyperparameter tuning, then evaluate on the remaining held-out instructions plus MATH and GSM8K; if the length-controlled win-rate margin over KABB shrinks to noise or reverses, the cognitive-alignment explanation for OSC's advantage is not supported.

Watch

Extended reading notes

Core claim

OSC claims that the way expert agents communicate, rather than the choice of experts or the aggregation of their answers, is what limits deep multi-agent LLM collaboration. To address this, each agent maintains a Collaborator Knowledge Model (CKM): a learned latent vector, updated by a GRU as dialogue proceeds, representing what that agent believes another agent knows, believes, or misunderstands about the task. A learnable gap function compares the agent's own cognitive state with the CKM-derived state of each collaborator, and a Proximal Policy Optimization (PPO) trained policy picks the communication action: which collaborator to address, which gap to close, what objective to pursue, and

Load-bearing premise

The result stands or falls on the evaluation being a fair, out-of-sample test of OSC's learned communication layer: the reported 81.4% win rate is measured on AlpacaEval 2.0 after tuning on its development subset, and the promised validation on complex reasoning benchmarks is not present in the reported experiments.

Editorial extensions

If this is right

  • If OSC's mechanism works as claimed, teams of off-the-shelf LLM experts can be made to collaborate more effectively by inserting a learned communication layer, without redesigning expert selection or aggregation.
  • Adaptive, gap-targeted messaging allows the same task quality with fewer dialogue rounds and fewer tokens, lowering inference cost at a fixed performance level.
  • Because the CKM, gap function, and policy are trained end-to-end from a composite reward, the same machinery can be optimized for other objectives, including explicit cost constraints or communication budgets.
  • The reported optimal team size is 6 agents; with 8 or 10 agents, coordination overhead—more rounds, more tokens, lower conflict resolution—begins to erode the gains.
  • A single-model variant of OSC outperforms the same model without the collaboration layer, suggesting the benefit does not depend on combining different model families.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The headline comparison is partly in-sample: the number of communication rounds and the communication-cost weight were tuned on a 160-instruction development subset of AlpacaEval 2.0, the same benchmark where the reported 81.4% win rate is measured; an independent hold-out test could yield a smaller margin.
  • The abstract says OSC was validated on complex reasoning and problem-solving benchmarks, but the reported experiments use AlpacaEval 2.0 and MT-Bench; MATH appears only as a stated training environment. Testing OSC on a clean reasoning benchmark such as MATH or GSM8K with no overlap between training and test tasks is the natural next check.
  • The same learned communication mechanism could transfer beyond LLM agents—to agents that are code modules, robots, or other learned communicators—whenever a state model and a language-realization layer exist.
  • The CKM's latent vectors could be probed against human judgments of confusion or confidence; if they do not track those states, OSC's advantage may come from extra computation or more conversational turns rather than from genuine cognitive modeling.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes OSC, a multi-agent LLM collaboration framework with trainable Collaborator Knowledge Models (CKM), a learnable cognitive-gap function fgap, and a reinforcement-learned communication policy pi_comm. The authors claim that OSC significantly improves task performance and communication efficiency, reporting an 81.4% LC win rate on AlpacaEval 2.0, a 9.94 average on MT-Bench, and a Pareto-optimal cost/performance curve. The abstract and introduction also claim validation on complex reasoning benchmarks such as MATH. The main evidence is Tables 1-4 and several appendix experiments, with the headline result trained via PPO on a subset of the evaluation benchmark.

Significance. If the central claims were cleanly supported, this would be a useful contribution: it addresses an underexplored intermediate layer between expert selection and answer aggregation, and its components are explicitly trainable with RL, moving beyond static role- or debate-based communication. The idea of modeling collaborators' cognitive states and learning communication objectives is plausible and potentially impactful. However, the paper's current evidence does not support the advertised claims: the headline evaluation is contaminated by training on the same benchmark, the promised reasoning-benchmark results are absent, and no uncertainty quantification is provided for the reported improvements. These are fixable in a revision, but they are load-bearing.

major comments (3)
  1. [§4.4, §5, §10, Table 1] The headline 81.4% LC win rate is not a clean out-of-sample comparison. §4.4 states that AlpacaEval 2.0 uses 805 instructions for training and evaluation with 160 reserved for development and 160 for validation; §5 says fine-tuning uses 160 of the 805 instructions; §10 confirms that the development set is used for training and the validation set for evaluation. Thus the OSC result reported in Table 1 appears to be measured on a 160-instruction validation subset, while the baseline values (KABB, MoA, DeepSeek-R1, etc.) are either public-leaderboard numbers on the full 805-instruction set or reproduced on the full set. This makes the comparison invalid. The authors must either evaluate on the full 805-instruction set without training on it, or evaluate all baselines on exactly the same held-out subset.
  2. [Abstract, §1, §4.1, §7.6] The abstract and §1 claim validation on 'complex reasoning and problem-solving benchmarks' and specifically MATH, and §1 lists MATH as a contribution. §7.6 states that training environments are constructed from MATH and GSM8K. However, no MATH, GSM8K, or any other reasoning benchmark result appears anywhere in the paper. The only evaluations reported are AlpacaEval 2.0 (instruction following) and MT-Bench (chat). The authors should either provide the missing reasoning-benchmark results or substantially revise the claims to match the actual experiments.
  3. [Tables 1-4] The paper repeatedly uses the word 'significantly' (e.g., §1, §4.1, §4.3) but reports no error bars, confidence intervals, or significance tests. §4.4 says results are averaged over 3 runs, yet Table 4 contains only point estimates. The differences between OSC and KABB on AlpacaEval (81.4 vs. 77.9) and between OSC and DeepSeek-R1 (81.4 vs. 80.1) are small; without variance or statistical testing, the claimed significance is unsupported. Add intervals or tests, at least for the headline metric.
minor comments (5)
  1. [Table 2] The baselines TalkHier, REMALIS, DyLAN, and MAC are not defined or referenced. Their experimental setup, model pool, and metric definitions are missing, so the communication-efficiency comparison cannot be reproduced or assessed.
  2. [§4.6] The phrase 'To address reviewer requests 6' should be removed; it is not appropriate in a submitted manuscript.
  3. [§4.1, §4.2] There are incomplete sentences. For example, §4.1 ends with 'Additional assessments include MT-Bench(...)' and §4.2 contains the repeated heading 'Experimental Setup Experimental Results' with no setup description.
  4. [§5, §9] Stray footnote markers or placeholder characters appear as '4' at the end of sentences (e.g., '...dynamic collaboration benefits 4.' and '...reasonable rounds and token count 4.'). These should be cleaned up.
  5. [§7.3] The reference to 'Equation 6 in Section 3.4' is off: the state representation is Eq. (5) in Section 3.4. Equation (6) is the message-generation equation.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline AlpacaEval number is an in-sample fit: OSC is PPO-trained on AlpacaEval 2.0 and evaluated on a 160-instruction subset of the same benchmark; reported reasoning-benchmark validation is absent.

  1. fitted input called prediction [Section 5 (Pretraining and Fine-tuning) and Section 10 (Reward Function Component Analysis)]
    "Fine-tuning: On AlpacaEval 2.0 (805 instructions, 160 for fine-tuning, 160 for validation) using PPO, 5×10^6 steps, reward R=R_task −0.001·C_comm + 0.05. ... The dataset used is AlpacaEval 2.0 (containing 805 instructions), with its development set (approx. 160 instructions) used for training and the validation set (approx. 160 instructions) for evaluation."

    The headline 81.4% LC win rate is not an independent prediction; it is the value of the PPO objective on the benchmark the system was trained on. Section 5 states that fine-tuning uses 160 AlpacaEval instructions and validation uses another 160 instructions from the same dataset. Section 10 confirms that the development set is used for training and the validation set for evaluation. Thus the reported improvement over KABB and MoA is the optimized training objective on a subset of the same distribution, not a held-out generalization result. The comparison is additionally broken because Table 1 labels the score as 'AlpacaEval 2.0 (805 instructions)' while the appendix clarifies evaluation is on only the 160-instruction validation split; leaderboard baselines were assessed on the full 805 ins

full rationale

The central scientific claim—that OSC significantly improves task performance and communication efficiency—is not demonstrated out-of-sample. Section 5 and Section 10 disclose that the system was fine-tuned with PPO on a 160-instruction development subset of AlpacaEval 2.0 and then evaluated on the 160-instruction validation subset of the same benchmark; Section 9 additionally describes hyperparameter tuning on that development set. The reported 81.4% LC win rate is therefore the optimized value of the training objective on the same distribution, and Table 1's comparison to baselines evaluated on the full 805-instruction leaderboard is invalid. This is a fitted-input-called-prediction circularity for the empirical claim, even though the algorithmic components themselves (CKM, fgap, πcomm) are not definitionally circular. I also flag, as a separate correctness issue rather than circularity, that the abstract and Contribution 3 claim validation on 'complex reasoning benchmarks (MATH)', but no MATH results appear in any table; only AlpacaEval 2.0 and MT-Bench are reported. The self-citation to KABB (Zhang et al. 2025d, same author group) is used only as a baseline and a source of the expert pool, not as load-bearing evidence for OSC's mechanisms, so it does not add to the circularity score.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The framework introduces multiple learned modules that are all fitted to the evaluation benchmark. The task reward, shaped reward, and communication cost are all defined on the AlpacaEval 2.0 benchmark family, and the reported performance is in-sample. The axioms are domain assumptions about the validity of the evaluation metrics and the reliability of RL optimization.

free parameters (4)
  • Nround = 4
    Number of communication rounds; tuned via grid search on the AlpacaEval 2.0 development set (Section 9).
  • lambda_cost = 0.001
    Communication cost weight in the reward; tuned on the development set (Section 9).
  • rshape = 0.05
    Intrinsic shaping reward, hand-set and tested in reward component analysis (Appendix 10).
  • dckm, dgru, dgap = 128, 128, 64
    Architectural dimensions, chosen without justification and varied in fine-grained ablation.
assumptions (4)
  • domain assumption AlpacaEval 2.0 LC win rate is a valid proxy for task performance and cognitive synergy.
    The entire evaluation rests on this metric, which measures instruction-following win rate against GPT-4, not deep reasoning.
  • domain assumption GPT-4-based evaluator reliably judges response quality.
    The paper uses AlpacaEval's default evaluator without validating its reliability for this multi-agent setting.
  • domain assumption The communication metrics (redundancy, conflict resolution, info density) are measurable and meaningful.
    No precise definitions or measurement procedures are given for these metrics; they seem to be asserted.
  • domain assumption PPO can effectively optimize the communication policy in the high-dimensional abstract action space.
    The paper assumes that RL will find good communication strategies, but no learning curves or convergence analysis are shown.
invented entities (3)
  • Collaborator Knowledge Model (CKM)
    purpose: A learned neural module that models each collaborator's cognitive state.
    No falsifiable prediction outside the paper; its internal representations are not externally validated.
  • Cognitive gap function fgap
    purpose: Learned module to quantify divergence between an agent's state and its model of a collaborator.
    No independent evidence that these gaps correspond to real communication needs.
  • Adaptive communication policy pi_comm
    purpose: RL policy that selects abstract communication actions.
    Trained to optimize the benchmark-specific reward; no evidence of generalizable communication strategies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OSC: Cognitive Orchestration through Dynamic Knowledge Alignment in Multi-Agent LLM Collaboration." pith.science (2026). https://pith.science/paper/3K4KJDLW

@misc{pith2026250904876,
  author       = {Pith},
  title        = {Pith review of: OSC: Cognitive Orchestration through Dynamic Knowledge Alignment in Multi-Agent LLM Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3K4KJDLW}},
  note         = {Machine review of arXiv:2509.04876}
}
read the original abstract

This paper introduces OSC (Orchestrating Cognitive Synergy), a knowledge-aware adaptive collaboration framework designed to enhance cognitive synergy in multi-agent systems with large language models. While prior work has advanced agent selection and result aggregation, efficient linguistic interactions for deep collaboration among expert agents remain a critical bottleneck. OSC addresses this gap as a pivotal intermediate layer between selection and aggregation, introducing Collaborator Knowledge Models (CKM) to enable each agent to dynamically perceive its collaborators' cognitive states. Through real-time cognitive gap analysis, agents adaptively adjust communication behaviors, including content focus, detail level, and expression style, using learned strategies. Experiments on complex reasoning and problem-solving benchmarks demonstrate that OSC significantly improves task performance and communication efficiency, transforming "parallel-working individuals'' into a "deeply collaborative cognitive team.'' This framework not only optimizes multi-agent collaboration but also offers new insights into LLM agent interaction behaviors.

Figures

Figures reproduced from arXiv: 2509.04876 by the authors.

Figure 1
Figure 1. The OSC framework uses Collaborator Knowledge Models (CKMs) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Price-performance trade-off on AlpacaEval [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. A comparison showing that fine-tuning the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Hyperparameter tuning for communication rounds (Nround) on AlpacaEval 2.0 shows that Nround = 4 achieves the optimal balance between task success (LC Win Rate) and communication cost (Avg. Tokens) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 14 canonical work pages

  1. [1]

    Qwen2 technical report

    2024. Qwen2 technical report

    work page 2024

  2. [2]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  3. [3]

    Jake Brawer, Kayleigh Bishop, Bradley Hayes, and Alessandro Roncone. 2023. https://arxiv.org/abs/2311.00153 Towards a natural language interface for flexible multi-agent task assignment . Preprint, arXiv:2311.00153

    work page Pith review arXiv 2023

  4. [4]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, and 12 others. 2020. https://arxiv.org/abs/2005.14165 Lan...

    arXiv 2020

  5. [5]

    Lingjiao Chen, Jared Quincy Davis, Boris Hanin, Peter Bailis, Ion Stoica, Matei Zaharia, and James Zou. 2024 a . https://arxiv.org/abs/2403.02419 Are more llm calls all you need? towards scaling laws of compound inference systems . Preprint, arXiv:2403.02419

    arXiv 2024

  6. [6]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2024 b . Frugalgpt: How to use large language models while reducing cost and improving performance. Transactions on Machine Learning Research

    work page 2024

  7. [7]

    DeepSeek-AI. 2024. https://arxiv.org/abs/2412.19437 Deepseek-v3 technical report . Preprint, arXiv:2412.19437

    arXiv 2024

  8. [8]

    DeepSeek-AI. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . Preprint, arXiv:2501.12948

    arXiv 2025

Show all 43 references
  1. [9]

    Junnan Dong, Qinggang Zhang, Chuang Zhou, Hao Chen, Daochen Zha, and Xiao Huang. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/d0aafec03d59db29a92fa683bd783374-Paper-Conference.pdf Cost-efficient knowledge-based question answering with large language models ...

    2024

  2. [10]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2023. https://arxiv.org/abs/2305.14325 Improving factuality and reasoning in language models through multiagent debate . Preprint, arXiv:2305.14325

    2023 arXiv

  3. [11]

    Zhuoyun Du, Chen Qian, Wei Liu, Zihao Xie, Yifei Wang, Yufan Dang, Weize Chen, and Cheng Yang. 2024. https://arxiv.org/abs/2406.08979 Multi-agent software development through cross-team collaboration . Preprint, arXiv:2406.08979

    2024 arXiv

  4. [12]

    Hang Gao and Yongfeng Zhang. 2024. https://arxiv.org/abs/2404.09982 Memory sharing for large language model based agents . Preprint, arXiv:2404.09982

    2024

  5. [13]

    Strategies for team success in llm application development

    Joe El Khoury GenAI. Strategies for team success in llm application development. https://medium.com/@jelkhoury880

  6. [14]

    Natalia Graziuso, Andrea Zugarini, and Stefano Melacci. 2024. https://aclanthology.org/2024.clicit-1.49/ Task-incremental learning on long text sequences . In Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024), pages 410--416, Pisa, Italy. C...

    2024

  7. [15]

    Chawla, Olaf Wiest, and Xiangliang Zhang

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. https://doi.org/10.24963/ijcai.2024/890 Large language model based multi-agents: A survey of progress and challenges . In Proceedings of the Thirty-Third ...

    2024 doi

  8. [16]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. NeurIPS

    2021

  9. [17]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and J \"u rgen Schmidhuber. 2024. Meta GPT : Meta programming for a multi-agent coll...

    2024

  10. [18]

    Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. 2024. https://doi.org/10.18653/v1/2024.acl-long.696 Harder task needs more experts: Dynamic routing in M o E models . In Proceedings of the 62nd...

    2024 doi

  11. [19]

    Mladjan Jovanovic and Peter Voss. 2024. https://arxiv.org/abs/2404.18311 Towards incremental learning in large language models: A critical review . Preprint, arXiv:2404.18311

    2024 arXiv

  12. [20]

    Bowman, Tim Rockt\" a schel, and Ethan Perez

    Akbir Khan, John Hughes, Dan Valentine, Laura Ruis, Kshitij Sachan, Ansh Radhakrishnan, Edward Grefenstette, Samuel R. Bowman, Tim Rockt\" a schel, and Ethan Perez. 2024. Debating with more persuasive llms leads to more truthful answers. In Proceedings of the 41st Internationa...

    2024

  13. [21]

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023 a . Camel: Communicative agents for "mind" exploration of large language model society. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  14. [22]

    Hashimoto

    Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023 b . Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval

    2023

  15. [23]

    OpenAI. 2024. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2024 arXiv

  16. [24]

    Deepak Babu Piskala, Vijay Raajaa, Sachin Mishra, and Bruno Bozza. 2024. https://doi.org/10.5120/ijca2024924172 Optiroute dynamic llm routing and selection based on user preferences: Balancing performance, cost, and ethics . International Journal of Computer Applications, 186(51):1–7

    2024 doi

  17. [25]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners

    2019

  18. [26]

    Manish Sanwal. 2025. https://arxiv.org/abs/2501.18645 Layered chain-of-thought prompting for multi-agent llm systems: A comprehensive approach to explainable large language models . Preprint, arXiv:2501.18645

    2025 arXiv

  19. [27]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. https://arxiv.org/abs/1707.06347 Proximal policy optimization algorithms . Preprint, arXiv:1707.06347

    2017 arXiv

  20. [28]

    Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch

    Vighnesh Subramaniam, Yilun Du, Joshua B. Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch. 2025. https://arxiv.org/abs/2501.05707 Multiagent finetuning: Self improvement with diverse reasoning chains . Preprint, arXiv:2501.05707

    2025 arXiv

  21. [29]

    Gemma Team. 2024. https://arxiv.org/abs/2408.00118 Gemma 2: Improving open language models at a practical size . Preprint, arXiv:2408.00118

    2024 arXiv

  22. [30]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. https://arxiv.org/abs/2302.13971 Llama:...

    2023 arXiv

  23. [31]

    Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. 2024 a . https://arxiv.org/abs/2406.04692 Mixture-of-agents enhances large language model capabilities . Preprint, arXiv:2406.04692

    2024 arXiv

  24. [32]

    Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, and Yangqiu Song. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.331 Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key? In Proceedings of the 62nd Annual Meeting of the Association for Computational...

    2024 doi

  25. [33]

    Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. https://arxiv.org/abs/2304.12244 Wizardlm: Empowering large language models to follow complex instructions . Preprint, arXiv:2304.12244

    2023 arXiv

  26. [34]

    Luke Yoffe, Alfonso Amayuelas, and William Yang Wang. 2025. https://arxiv.org/abs/2407.06426 Debunc: Improving large language model agent communication with uncertainty metrics . Preprint, arXiv:2407.06426

    2025 arXiv

  27. [35]

    Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu, Bryan Hooi, and Shumin Deng. 2024 a . https://arxiv.org/abs/2310.02124 Exploring collaboration mechanisms for llm agents: A social psychology view . Preprint, arXiv:2310.02124

    2024 arXiv

  28. [36]

    Jusheng Zhang, Kaitong Cai, Yijia Fan, Jian Wang, and Keze Wang. 2025 a . https://arxiv.org/abs/2506.17267 Cf-vlm:counterfactual vision-language fine-tuning . Preprint, arXiv:2506.17267

    2025 arXiv

  29. [37]

    Jusheng Zhang, Yijia Fan, Kaitong Cai, and Keze Wang. 2025 b . https://arxiv.org/abs/2502.06018 Kolmogorov-arnold fourier networks . Preprint, arXiv:2502.06018

    2025 arXiv

  30. [38]

    Jusheng Zhang, Yijia Fan, Wenjun Lin, Ruiqi Chen, Haoyi Jiang, Wenhao Chai, Jian Wang, and Keze Wang. 2025 c . Gam-agent: Game-theoretic and uncertainty-aware collaboration for complex visual reasoning. arXiv preprint arXiv:2505.23399

    2025 arXiv

  31. [39]

    Jusheng Zhang, Zimeng Huang, Yijia Fan, Ningyuan Liu, Mingyan Li, Zhuojie Yang, Jiawei Yao, Jian Wang, and Keze Wang. 2025 d . https://openreview.net/forum?id=AKvy9a4jho KABB : Knowledge-aware bayesian bandits for dynamic expert coordination in multi-agent systems . In Forty-s...

    2025

  32. [40]

    Yi Zhang, Sen Wang, Zhi Chen, Xuwei Xu, Stano Funiak, and Jiajun Liu. 2024 b . https://doi.org/10.1007/978-981-97-2253-2_14 Towards cost-efficient federated multi-agent rl with learnable aggregation . page 171–183, Berlin, Heidelberg. Springer-Verlag

    2024 doi

  33. [41]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. In Proceedings of the 37th Internat...

    2023

  34. [42]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  35. [43]

    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 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.