Pith. sign in

REVIEW 3 major objections 6 minor 30 references

Towards Cognitive Synergy in LLM-Based Multi-Agent Systems: Integrating Theory of Mind and Critical Evaluation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding prompt-based theory of mind and a structured critic agent to an LLM-based multi-agent system produces emergent cognitive synergy, evidenced by the combined configuration's highest scores on argument quality…

desk verdict Honest pilot combining prompted ToM and an LLM critic, but the termination confound breaks the synergy claim; salvageable with proper controls. read the letter →

arxiv 2507.21969 v1 pith:BGRESVXR submitted 2025-07-29 cs.MA

Adam Kostka , Jaros{l}aw A. Chudziak This is my paper
classification cs.MA
keywords multi-agentsystemstheoryofmindcriticalevaluationcognitivesynergylargelanguagemodelsLLM-basedagentspromptengineeringriskresolution
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 sets out to test whether two human-like cognitive mechanisms—anticipating what other agents will say (theory of mind) and systematically attacking arguments (a structured critic)—can make a team of LLM agents reason better than they do with either mechanism alone or with neither. In a strategic investment case study, the configuration with both mechanisms scored highest on argument quality, critical engagement, and risk resolution, and produced the most revision triggers. The authors interpret this as evidence of emergent cognitive synergy, where the combined system is more than the sum of its parts. The claim matters because multi-agent LLM systems currently coordinate poorly and rarely challenge their own reasoning, so a cheap prompting-based route to more rigorous collective reasoning would be broadly useful.

What carries the argument

The mechanism is a prompt-level implementation of theory of mind plus a separate Critic Agent, orchestrated by an Integrator and an Orchestrator over a knowledge graph with logical inference. When theory of mind is enabled, expert agents are instructed to first state what a named peer is likely to argue given its role, and then frame their own contribution to complement or counter that anticipated view. The Critic Agent, enabled after the first round, evaluates arguments for logical consistency, unsupported claims, and missed constraints without advancing its own solution. The Orchestrator decides when the discussion has reached resolution, which turns out to be the load-bearing control point in the experiment.

What would settle it

Run all four configurations with the Orchestrator's stopping rule fixed to a constant number of dialogue turns, or with dialogue length matched across conditions, and check whether the ToM plus Critic configuration still resolves roughly 49% of risks while the baseline resolves none. Alternatively, have expert human annotators score the same dialogues with the same rubric; if human scores do not reproduce the LLM judge's ranking, the reported gaps are an artifact of the judge rather than the mechanisms.

Watch

Extended reading notes

Core claim

The paper's central claim is that integrating dynamic theory-of-mind prompting with a dedicated Critic Agent produces emergent cognitive synergy in an LLM-based multi-agent system. In the best configuration (ToM plus Critic), an external LLM judge rated argument quality at 3.43 versus 2.75 for the baseline, critical engagement at 0.38 versus 0.00, and risk resolution at 48.75% versus 0.00%, with the most revision triggers per conversation (1.75). The authors argue that theory of mind improves coordination and referential cohesion while the critic deepens analysis and surfaces overlooked risks, and that the two mechanisms reinforce each other rather than simply adding up. The paper also notes that these results come from a single case study and depend on an LLM judge without direct human validation.

Load-bearing premise

The outcome differences across the four configurations are caused by the manipulated mechanisms (theory-of-mind prompts and the Critic's presence) rather than by how long the dialogues ran and when the Orchestrator stopped them.

Editorial extensions

If this is right

  • Configurations with theory of mind produce more cohesive references across turns, indicating better coordination and less redundant reasoning.
  • Configurations with the Critic Agent show higher argument quality, deeper feedback integration, and more revision triggers, consistent with the claim that structured critique sharpens reasoning.
  • The best combined configuration yields the highest risk resolution (48.75%) and the most adaptation, which the paper interprets as synergy beyond either mechanism alone.
  • Without a critic, the Orchestrator tends to terminate discussions prematurely, so the stopping rule is a key design lever for dialogue depth.
  • The knowledge-base and formal-reasoning tool played only a small role in this short dialogue, so its contribution is deferred to longer-horizon settings.

Reading between the lines

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

  • A testable extension: hold dialogue length constant across configurations to separate the cognitive mechanisms from the stopping-rule effect; the current design entangles them.
  • If the synergy is genuine, the same prompt-and-critic pattern could be dropped into negotiation, legal reasoning, and collaborative research settings, where iterative anticipation and critique are natural.
  • The LLM-judge dependence means the headline gaps should be re-measured with human annotation before being treated as evidence about cognition rather than about judge bias.
  • The 'emergent synergy' framing implies that the two mechanisms interact, but the current evidence supports only an additive score advantage; an interaction term or crossover design would test emergence more directly.
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 / 6 minor

Summary. The paper proposes an LLM-based multi-agent system that augments three expert agents with two cognitive mechanisms: a prompt-based Theory of Mind (ToM) module that instructs agents to anticipate peers' arguments, and a Critic Agent that performs structured critique. The system is evaluated in a single case study—a strategic R&D budget allocation decision—under four configurations (FF, TF, FT, TT) using an LLM judge with a custom rubric. The paper reports that the ToM+Critic configuration (TT) scores highest on argument quality, critical engagement, risk resolution, and revision triggers, and interprets this as evidence of 'emergent cognitive synergy' exceeding the sum of parts. The authors also note limitations including the lack of human validation of the LLM judge and premature dialogue termination in no-critic configurations.

Significance. If the empirical claims were robust, the paper would offer a simple, prompt-level recipe for improving multi-agent coordination and reasoning—valuable for practitioners building LLM-based MAS. The conceptual framing connecting ToM, structured critique, and collective intelligence is clear, and the four-condition factorial design is a reasonable starting point. However, the paper's significance is currently limited by the lack of statistical controls, a single case study, and an evaluation powered entirely by an unvalidated LLM judge. The authors deserve credit for explicitly acknowledging the premature-termination issue and the absence of human annotation, but those concessions effectively undermine the headline synergy claim unless addressed with additional experiments.

major comments (3)
  1. The central 'emergent cognitive synergy' claim is confounded by the Orchestrator's termination heuristic. The Discussion states, 'In both TF and FF, the Orchestrator Agent consistently interrupted the dialogue before all critical aspects were discussed.' Because risk resolution and revision triggers are measured over the dialogue (see Evaluation), truncated dialogues mechanically have fewer opportunities to surface or resolve risks and cannot generate revision triggers in the absence of a Critic Agent. Thus the headline TT-versus-FF gaps (48.75% vs 0.00% risk resolution; 1.75 vs 0.00 revision triggers) could arise entirely from the stopping policy rather than from any cognitive mechanism. The paper provides no control for dialogue length or stopping criterion across configurations, and Tables 1 and 2 report single numbers with no run count, variance, or significance test. The authors' own limitation statement confirms the issue, but no remedy is offered; this must be fixed with either a length-matched analysis, a fixed-turn protocol, or at minimum explicit reporting of turn counts per configuration.
  2. The evaluation protocol does not support the quantitative comparisons. The paper does not state how many conversations were run per configuration, whether multiple seeds or temperature settings were used, or whether the reported values are means, medians, or single-run outcomes. The sole outcome measure is an external LLM judge applying the authors' rubric, and the Discussion concedes that 'the results even lacked direct validation against expert human annotation for this specific task.' Given that the judge is a language model similar to the agents under test, the 3.43 vs 3.15 vs 3.00 differences in Table 1 may reflect judge noise rather than true differences. The manuscript should report the number of runs, per-run scores, inter-rater agreement (or at least human-judge correlation on a subset), and appropriate significance tests or effect sizes before claiming that 'TT performed best' or that 'neither module alone guaranteed high performance.'
  3. The 'exceeds the sum of its parts' synergy claim is not actually tested. A super-additive synergy claim requires a comparison against an additive model—for example, showing that TT's improvement over baseline exceeds the sum of the individual improvements of TF and FT, or an interaction effect in a factorial ANOVA. Table 1 shows TT at 3.43 against TF at 3.15 and FT at 3.00, which is merely ordinal superiority, not evidence of emergence. The paper should define a formal synergy criterion and test it, or temper the conclusion to claim additive or complementary benefits rather than emergent synergy.
minor comments (6)
  1. The prompt templates for ToM and Critic modules are described only in prose; providing the exact prompts in an appendix would make the intervention reproducible and would help readers assess whether the ToM manipulation is as strong as claimed.
  2. The manuscript never reports the number of dialogue turns per configuration in the results, even though the Discussion describes shorter dialogues for TF and FF; turn counts are essential for interpreting the risk-resolution and revision-trigger metrics.
  3. The rubric describes 'critical engagement' as the proportion of substantive counterarguments out of total turns and 'risk resolution' as the proportion of discovered risks addressed, but the denominator (number of risks/turns) is never reported; without these denominators the percentages in Table 1 cannot be interpreted.
  4. The citation to 'Pan, Albalak, Wang, & Wang, 2023' appears to refer to Logic-LM, but the paper does not clearly explain how Logic-LM relates to self-correction or critic mechanisms; please clarify the connection.
  5. The text contains minor grammatical issues (e.g., 'Clingo solver' and 'In case information retrieval is lacking') and inconsistent use of hyphens; a careful language edit is recommended.
  6. Figure 4 is mentioned in the Discussion but not described in the text; please ensure it is introduced and interpreted explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper is an empirical comparison, and its main threats are validity issues rather than input-output circularity.

full rationale

The paper contains no formal derivation chain in which a predicted quantity is recovered from its own definition or from a fitted input renamed as a result. The central claim—that the ToM + Critic configuration outperforms the baseline—is supported by observed dialogue-level metrics, not by an equation that reduces to its own assumptions. The self-citations present (Kostka & Chudziak 2024; Chudziak & Wawer 2024; Chudziak & Kostka 2025; Cinkusz & Chudziak 2024) are background and motivation; none is load-bearing for the quantitative claim, none invokes a uniqueness theorem, and none smuggles in a contested ansatz. The evaluation uses an LLM judge rather than human annotation, and the Discussion itself concedes that the Orchestrator 'consistently interrupted the dialogue before all critical aspects were discussed' in the no-critic conditions; these are genuine threats to experimental validity and generalizability, but they are not cases where a prediction is equivalent to its input by construction. No fitted parameter is disguised as a prediction, and no cited prior result is used to forbid alternatives. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

Design choices that function as tuning parameters: the Orchestrator's termination and reactivation heuristics, and the exact prompt templates that operationalize ToM and critique, neither released nor systematically varied. The evaluation additionally assumes that all outcome differences are attributable to the two mechanisms, which the paper's own Discussion contradicts via the premature-termination admission, and that the LLM judge's rubric scores validly measure reasoning quality, which the paper concedes was never validated against human annotation. The single invented conceptual entity, cognitive synergy, has no falsifiable handle beyond the paper's own rubric.

free parameters (2)
  • Orchestrator termination and reactivation heuristics = not specified
    Hand-authored rules decide when a dialogue ends ('repeated until a satisfactory level of resolution'); the authors report this rule terminated baseline dialogues prematurely, so it operates as an unmeasured setting that changes outcomes across conditions.
  • ToM and Critic prompt templates = not released
    The mechanisms are implemented purely by prompt text whose exact wording is paraphrased rather than quoted; in LLM systems the prompt wording is the implementation, and this hand-chosen text materially determines the results.
assumptions (3)
  • domain assumption Outcome differences are attributable to the two manipulated mechanisms alone
    The experimental contrast assumes no other systematic difference between conditions; the Discussion's admission of premature Orchestrator termination in FF and TF undercuts this assumption.
  • domain assumption LLM judge ratings validly measure argument quality, cohesion, and risk resolution
    The paper uses an external LLM-as-judge with a rubric and no human validation; the Discussion explicitly concedes the results 'lacked direct validation against expert human annotation for this specific task.'
  • domain assumption Prompted role anticipation is a faithful operationalization of Theory of Mind
    The Methodology states 'this methodology simulates the human act of mentalizing'; no independent measure verifies that agents model others' mental states rather than echoing role stereotypes.
invented entities (1)
  • Cognitive synergy (emergent super-additivity of agent capabilities)
    purpose: Frames the combined ToM plus Critic result as the collective exceeding the sum of its parts; it is the central conceptual contribution claimed in the abstract.
    No falsifiable handle outside the paper's own rubric: no test distinguishes super-additivity from additive or confounded effects (longer dialogues with critique turns naturally resolve more risks), so the entity carries the strongest claim without independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Cognitive Synergy in LLM-Based Multi-Agent Systems: Integrating Theory of Mind and Critical Evaluation." pith.science (2026). https://pith.science/paper/BGRESVXR

@misc{pith2026250721969,
  author       = {Pith},
  title        = {Pith review of: Towards Cognitive Synergy in LLM-Based Multi-Agent Systems: Integrating Theory of Mind and Critical Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGRESVXR}},
  note         = {Machine review of arXiv:2507.21969}
}
read the original abstract

Recently, the field of Multi-Agent Systems (MAS) has gained popularity as researchers are trying to develop artificial intelligence capable of efficient collective reasoning. Agents based on Large Language Models (LLMs) perform well in isolated tasks, yet struggle with higher-order cognition required for adaptive collaboration. Human teams achieve synergy not only through knowledge sharing, but also through recursive reasoning, structured critique, and the ability to infer others' mental states. Current artificial systems lack these essential mechanisms, limiting their ability to engage in sophisticated collective reasoning. This work explores cognitive processes that enable effective collaboration, focusing on adaptive theory of mind (ToM) and systematic critical evaluation. We investigate three key questions. First, how does the ability to model others' perspectives enhance coordination and reduce redundant reasoning? Second, to what extent does structured critique improve reasoning quality by identifying logical gaps and mitigating biases? Third, the interplay of these mechanisms can lead to emergent cognitive synergy, where the collective intelligence of the system exceeds the sum of its parts. Through an empirical case study on complex decision making, we show that the integration of these cognitive mechanisms leads to more coherent, adaptive, and rigorous agent interactions. This article contributes to the field of cognitive science and AI research by presenting a structured framework that emulates human-like collaborative reasoning MAS. It highlights the significance of dynamic ToM and critical evaluation in advancing multi-agent systems' ability to tackle complex, real-world challenges.

Figures

Figures reproduced from arXiv: 2507.21969 by the authors.

Figure 1
Figure 1. Conceptual Cycle of Cognitive Synergy in Multi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustrative Comparison of Hypothesized Agent Re [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. System architecture showing the interaction among [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Multi-Round Analysis Process with Critic-Driven [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 11 canonical work pages

  1. [1]

    write newline

    " write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot := field num.names 'numnames := numnames 'format.num.names := format.num.names na...

  2. [2]

    \ Chudziak, J A

    Choiski2009OntologicalLA APACrefauthors Choi \'n ski, M. \ Chudziak, J A. APACrefauthors \ 2009 . Ontological learning assistant for knowledge discovery and data mining Ontological learning assistant for knowledge discovery and data mining . 2009 international multiconference on computer science and information technology 2009 international multiconferenc...

    work page 2009

  3. [3]

    \ Kostka, A

    ChudziakKostkaAIED2025 APACrefauthors Chudziak, J A. \ Kostka, A. APACrefauthors \ 2025 . AI-powered math tutoring: Platform for personalized and adaptive education AI-powered math tutoring: Platform for personalized and adaptive education . Proceedings of the 26th international conference on artificial intelligence in education ( AIED 2025). Proceedings ...

    work page 2025

  4. [4]

    \ Wawer, M

    paclic_elliottagents APACrefauthors Chudziak, J A. \ Wawer, M. APACrefauthors \ 2024 . ElliottAgents : A natural language-driven multi-agent system for stock market analysis and prediction ElliottAgents : A natural language-driven multi-agent system for stock market analysis and prediction . Proceedings of the 38th pacific asia conference on language, inf...

    work page 2024

  5. [5]

    \ Chudziak, J A

    10.1145/3691620.3695336 APACrefauthors Cinkusz, K. \ Chudziak, J A. APACrefauthors \ 2024 . Towards LLM -augmented multiagent systems for agile software engineering Towards LLM -augmented multiagent systems for agile software engineering . Proceedings of the 39th IEEE / ACM international conference on automated software engineering Proceedings of the 39th...

    arXiv 2024

  6. [6]

    , Xiang, V

    cross2024hypothetical APACrefauthors Cross, L. , Xiang, V. , Bhatia, A. , Yamins, D L. \ Haber, N. APACrefauthors \ 2024 . Hypothetical minds: Scaffolding theory of mind for multi-agent tasks with large language models Hypothetical minds: Scaffolding theory of mind for multi-agent tasks with large language models . NeurIPS 2024 workshop on open-world agen...

    work page 2024

  7. [7]

    \ Mesmer-Magnus, J

    dech2010 APACrefauthors Dechurch, L. \ Mesmer-Magnus, J. APACrefauthors \ 2010 . The cognitive underpinnings of effective teamwork: A meta-analysis The cognitive underpinnings of effective teamwork: A meta-analysis . The Journal of Applied Psychology 95 32--53 . APACrefDOI doi:10.1037/a0017328 APACrefDOI

    work page doi:10.1037/a0017328 2010

  8. [8]

    , Shao, Z

    DBLP:conf/iclr/GouSGSYDC24 APACrefauthors Gou, Z. , Shao, Z. , Gong, Y. , Shen, Y. , Yang, Y. , Duan, N. \ Chen, W. APACrefauthors \ 2024 . CRITIC : Large language models can self-correct with tool-interactive critiquing CRITIC : Large language models can self-correct with tool-interactive critiquing . ICLR . ICLR . APACrefURL https://openreview.net/forum...

    work page 2024

Show all 30 references
  1. [9]

    , Jiang, X

    gu2025surveyllmasajudge APACrefauthors Gu, J. , Jiang, X. , Shi, Z. , Tan, H. , Zhai, X. , Xu, C. Guo, J. APACrefauthors \ 2025 . A survey on LLM -as-a- Judge . A survey on LLM -as-a- Judge . APACrefURL https://arxiv.org/abs/2411.15594 APACrefURL

    2025 arXiv

  2. [10]

    , Sycara, K

    jennings1998roadmap APACrefauthors Jennings, N. , Sycara, K. \ Wooldridge, M. APACrefauthors \ 1998 . A roadmap of agent research and development A roadmap of agent research and development . Autonomous Agents and Multi-Agent Systems 1 7--38 . APACrefDOI doi:10.1023/A:10100904...

    1998 doi

  3. [11]

    \ Chudziak, J A

    synergymas APACrefauthors Kostka, A. \ Chudziak, J A. APACrefauthors \ 2024 . Synergizing logical reasoning, long-term memory, and collaborative intelligence in multi-agent LLM systems Synergizing logical reasoning, long-term memory, and collaborative intelligence in multi-age...

    2024

  4. [12]

    , Xia, C S

    lee2024unifieddebuggingapproachllmbased APACrefauthors Lee, C. , Xia, C S. , Yang, L. , tse Huang, J. , Zhu, Z. , Zhang, L. \ Lyu, M R. APACrefauthors \ 2024 . A unified debugging approach via LLM -based multi-agent synergy . A unified debugging approach via LLM -based multi-a...

    2024

  5. [13]

    , Tan, Z

    li2024smoaimprovingmultiagentlarge APACrefauthors Li, D. , Tan, Z. , Qian, P. , Li, Y. , Chaudhary, K S. , Hu, L. \ Shen, J. APACrefauthors \ 2024 . SMoA : Improving multi-agent large language models with sparse mixture-of-agents . SMoA : Improving multi-agent large language m...

    2024 arXiv

  6. [14]

    , Chong, Y

    li-etal-2023-theory APACrefauthors Li, H. , Chong, Y. , Stepputtis, S. , Campbell, J. , Hughes, D. , Lewis, C. \ Sycara, K. APACrefauthors \ 2023 12 . Theory of mind for multi-agent collaboration via large language models Theory of mind for multi-agent collaboration via large ...

    2023

  7. [15]

    , Menon, V

    li2024challengesfacedlargelanguage APACrefauthors Li, P. , Menon, V. , Gudiguntla, B. , Ting, D. \ Zhou, L. APACrefauthors \ 2024 . Challenges faced by large language models in solving multi-agent flocking. Challenges faced by large language models in solving multi-agent flock...

    2024 arXiv

  8. [16]

    , Wang, S

    Li2024 APACrefauthors Li, X. , Wang, S. , Zeng, S. , Wu, Y. \ Yang, Y. APACrefauthors \ 2024 . A survey on LLM -based multi-agent systems: workflow, infrastructure, and challenges A survey on LLM -based multi-agent systems: workflow, infrastructure, and challenges . Vicinagear...

    2024 doi

  9. [17]

    APACrefauthors \ 2016

    MERCIER2016689 APACrefauthors Mercier, H. APACrefauthors \ 2016 . The argumentative theory: Predictions and empirical evidence The argumentative theory: Predictions and empirical evidence . Trends in Cognitive Sciences 20 9 689--700 . APACrefDOI doi:10.1016/j.tics.2016.07.001 ...

    2016 doi

  10. [18]

    \ Sperber, D

    MercierSperber2011 APACrefauthors Mercier, H. \ Sperber, D. APACrefauthors \ 2011 . Why do humans reason? Arguments for an argumentative theory Why do humans reason? arguments for an argumentative theory . Behavioral and Brain Sciences 34 2 57--74 . APACrefDOI doi:10.1017/S014...

    2011 doi

  11. [19]

    , Rizk, Y

    muthusamy-etal-2023-towards APACrefauthors Muthusamy, V. , Rizk, Y. , Kate, K. , Venkateswaran, P. , Isahagian, V. , Gulati, A. \ Dube, P. APACrefauthors \ 2023 12 . Towards large language model-based personal agents in the enterprise: Current trends and open problems Towards ...

    2023

  12. [20]

    , Khan, A U

    naveed2024comprehensiveoverviewlargelanguage APACrefauthors Naveed, H. , Khan, A U. , Qiu, S. , Saqib, M. , Anwar, S. , Usman, M. Mian, A. APACrefauthors \ 2024 . A comprehensive overview of large language models. A comprehensive overview of large language models. APACrefURL h...

    2024 arXiv

  13. [21]

    , Albalak, A

    pan-etal-2023-logic APACrefauthors Pan, L. , Albalak, A. , Wang, X. \ Wang, W. APACrefauthors \ 2023 12 . Logic- LM : Empowering large language models with symbolic solvers for faithful logical reasoning Logic- LM : Empowering large language models with symbolic solvers for fa...

    2023

  14. [22]

    \ Woodruff, G

    PremackWoodruff1978 APACrefauthors Premack, D. \ Woodruff, G. APACrefauthors \ 1978 . Does the chimpanzee have a theory of mind? Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences 1 4 515--526 . APACrefDOI doi:10.1017/S0140525X00076512 APACrefDOI

    1978 doi

  15. [23]

    , Wang, H

    qiu2024collaborativeintelligencepropagatingintentions APACrefauthors Qiu, X. , Wang, H. , Tan, X. , Qu, C. , Xiong, Y. , Cheng, Y. Qi, Y. APACrefauthors \ 2024 . Towards collaborative intelligence: Propagating intentions and reasoning for multi-agent coordination with large la...

    2024 arXiv

  16. [24]

    xi2023rise APACrefauthors Schmidgall, S. , Su, Y. , Wang, Z. , Sun, X. , Wu, J. , Yu, X. Barsoum, E. APACrefauthors \ 2025 . Agent laboratory: Using LLM agents as research assistants . Agent laboratory: Using LLM agents as research assistants . APACrefURL https://arxiv.org/abs...

    2025 arXiv

  17. [25]

    , Carpenter, M

    tom2005 APACrefauthors Tomasello, M. , Carpenter, M. , Call, J. , Behne, T. \ Moll, H. APACrefauthors \ 2005 . Understanding and sharing intentions: The origins of cultural cognition Understanding and sharing intentions: The origins of cultural cognition . Behavioral and Brain...

    2005 doi

  18. [26]

    , Wang, J

    wang2024mixtureofagentsenhanceslargelanguage APACrefauthors Wang, J. , Wang, J. , Athiwaratkun, B. , Zhang, C. \ Zou, J. APACrefauthors \ 2024 . Mixture-of-agents enhances large language model capabilities. Mixture-of-agents enhances large language model capabilities. APACrefU...

    2024 arXiv

  19. [27]

    , Carey, S

    Wellman1990 APACrefauthors Wellman, H M. , Carey, S. , Gleitman, L. , Newport, E L. \ Spelke, E S. APACrefauthors \ 1990 . The child's theory of mind The child's theory of mind . The MIT Press . APACrefDOI doi:10.7551/mitpress/1811.001.0001 APACrefDOI

    1990 doi

  20. [28]

    APACrefauthors \ 2009

    Wooldridge_2009 APACrefauthors Wooldridge, M J. APACrefauthors \ 2009 . An introduction to multiagent systems An introduction to multiagent systems . John Wiley & Sons

    2009

  21. [29]

    , Chabris, C F

    WoolleyEvidenceFA APACrefauthors Woolley, A W. , Chabris, C F. , Pentland, A. , Hashmi, N. \ Malone, T W. APACrefauthors \ 2010 . Evidence for a collective intelligence factor in the performance of human groups Evidence for a collective intelligence factor in the performance o...

    2010 doi

  22. [30]

    , Chiang, W L

    10.5555/3666122.3668142 APACrefauthors Zheng, L. , Chiang, W L. , Sheng, Y. , Zhuang, S. , Wu, Z. , Zhuang, Y. Stoica, I. APACrefauthors \ 2023 . Judging LLM -as-a-judge with MT -bench and Chatbot Arena Judging LLM -as-a-judge with MT -bench and Chatbot Arena . Proceedings of ...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.