Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Social Networks of LLM Agents

T0 review · 3 major / 5 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read A single attention-width knob decides whether LLM agent populations pool knowledge or herd into false consensus.

desk verdict Clean operator-level control of herding vs. wisdom in multi-agent LLMs, with real theorems and large ablated accuracy swings; the bridge-to-proxy is only non-vacuous on the anchored testbed, not on the headline DeGroot plots. read the letter →

arxiv 2607.03695 v1 pith:QXIJJQFV submitted 2026-07-04 cs.LG

Kaixuan Liu , Guojun Xiong , Weinan Zhang , Shengpu Tang This is my paper
classification cs.LG
keywords LLMmulti-agentsystemssociallearningopiniondynamicsattentionwidthherdingwisdomofcrowdsrealizedinfluenceequalizedallocation
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

When large language model agents talk to one another, what the group ends up believing can be either wiser than any member or confidently wrong. Classical social-network theory treats the visible links among agents as the whole story of how beliefs combine. That picture fails for language-model agents: each one has a limited context window and selective attention, so it only realizes part of what it is exposed to. The paper introduces SNLA, a framework that separates visible exposure from realized influence. Realized influence is built from each source’s social standing and a single temperature parameter that controls how sharply attention focuses. On a tractable mathematical proxy the authors prove that narrow attention collapses the effective sample size of the group to a constant independent of population size (herding), while wide attention recovers wisdom-of-crowds accuracy only when the exposure graph is balanced. A simple decentralized pricing rule that equalizes attention across sources restores the optimal collective weights. Controlled experiments and three multi-agent benchmarks show the same herding-to-wisdom transition, and the pricing fix removes herding without retraining the models.

What carries the argument

The realized-influence operator C: each reader reweights its exposure by source social power and then applies a temperature-β softmax, so that C, not the visible network, drives belief updates. A path-wise bridge theorem couples real LLM emissions to an anchored Friedkin–Johnsen proxy, after which two-regime and equalization theorems characterize effective sample size as a function of β and column balance.

What would settle it

On an aggregation-dependent multi-agent task with a fixed dominant wrong source, measure collective accuracy while sweeping attention width: if accuracy stays flat instead of rising sharply from a low floor at narrow width to near-oracle levels at wide width, or if equalizing column sums fails to lift the narrow-width floor, the central claim fails.

Watch

Extended reading notes

Core claim

The paper establishes that collective accuracy of an LLM-agent population is controlled by attention width: narrow attention produces herding whose effective sample size stays bounded no matter how large the population grows, while wide attention recovers wisdom-of-crowds behavior only on undirected degree-regular exposure graphs; a decentralized equalizer that drives the influence matrix toward double stochasticity restores optimal collective weights whenever a dominant source is present.

Load-bearing premise

The mathematical link from real language-model replies to the analyzable proxy requires every reply to stay close to an anchored update whose self-weight is strictly less than one; the main discussion benchmarks run without that anchor.

Editorial extensions

If this is right

  • Designers of multi-agent LLM systems can treat attention width as a single control that moves a population between herding and wisdom without changing model weights.
  • When a high-power source is present, one Sinkhorn-style price update per round is enough to keep collective variance near the optimal 1/n floor.
  • Placement of capable agents at high-degree nodes improves coordination even when there is no single wrong answer to herd onto.
  • Classical wisdom-of-crowds guarantees transfer to LLM societies only after exposure is made doubly stochastic and attention is sufficiently wide.

Reading between the lines

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

  • The same column-collapse pathology appears in mixture-of-experts routing; the paper’s equalizer is formally the same family of load-balancing fixes, suggesting a shared fix across agent societies and expert routing.
  • If the bridge can be made non-vacuous for unanchored DeGroot-style debate, the theory would directly certify the accuracy curves already plotted on the main benchmarks rather than only on the anchored testbed.
  • Operator-controlled variants of existing multi-agent suites could become a standard stress test for whether a new agent architecture herds or pools.
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 introduces SNLA, a social-network model for populations of LLM agents that separates visible exposure from realized influence. Realized influence is derived from exposure weights, Katz–Bonacich social power, and a finite attention width β via a tempered (temperature-softmax) allocation. On a tractable Friedkin–Johnsen proxy, the authors prove a path-wise bridge to δ-approximate LLM emissions under anchoring (Theorem 4.1), a two-regime result in which narrow β yields herding with Neff bounded independently of n while wide β recovers wisdom-of-crowds accuracy only for undirected degree-regular exposure (Theorem 4.2), and an equalization theorem showing that a decentralized Sinkhorn-style pricing protocol drives collective weights to uniform when column defects vanish (Theorem 4.3). Empirically, a controlled scalar testbed validates Neff/variance predictions and the bridge; operator-controlled variants of HiddenBench, Werewolf, and AgentsNet show large accuracy/coordination swings with β, placement effects, and equalizer recovery of herding.

Significance. If the results hold, the paper supplies a usable control (attention width β, plus an equalizer) for when multi-agent LLM systems pool information versus herd—directly relevant to debate, simulation, and collaborative agent systems. Strengths include explicit, self-contained proofs with stated assumptions; a measurable residual δ that makes the bridge falsifiable; a controlled testbed that reports Neff and the bridge bound (254/254 cells); large, ablated accuracy swings on recognized benchmarks; and cross-family replication on Llama. The MoE-routing analogy and the pricing protocol as information projection are additional conceptual contributions. The work is a genuine advance over classical DeGroot-style models that treat exposure as influence.

major comments (3)
  1. [§5.2, Theorem 4.1, Fig. 2A–B] §5.2 and Theorem 4.1 / Proposition 8.4: The primary collective-accuracy claims (Fig. 2A,B; Tables 5–9) use DeGroot updating (λ=1), where the authors correctly note the bridge bound is vacuous and that the discussion benchmarks share the β-gating mechanism rather than a theorem mapping Neff onto accuracy. The non-vacuous bridge evidence (Fig. 4; OLS slope 0.88, R²=0.74) lives on the anchored testbed (median fitted λ≈0.25). This is load-bearing for the packaged claim that the herding–wisdom transition “reproduces” the theory: please either (i) report proxy Neff or q_T diagnostics on the discussion runs themselves, or (ii) reframe the headline empirical claim more sharply as a mechanistic demonstration of the attention bottleneck, with theorem-backed Neff transfer reserved for the anchored testbed.
  2. [Appendix 7, §5.2, Fig. 4] Appendix 7 / §5.2 (bridge certification): Anchoring weights λ_i are least-squares fit from the same residuals ˆδ_i(t) that enter the uniform residual δ_T used in the bridge bound. After per-agent best-fit λ, “0 violations on 254 cells” is a statement about best-fit ceilings, not about a prescribed anchoring regime. The manuscript already notes this; please make the distinction fully explicit in the main text (not only the appendix) and report the distribution of fitted λ and of δ_T/(1−λ_max) separately for anchored-persona vs. unanchored cells so readers can judge how non-vacuous the bound is on the subset that actually supports Theorem 4.1.
  3. [Theorem 4.2, §9.3] Theorem 4.2 (narrow regime): The Neff≤8 bound relies on a unique dominant pair (j★_i = h for all i≠h, j★_h = g≠h) and ρ(β)<1. This is a strong structural assumption. Please state how often the unique-dominant-pair condition holds on the exposure graphs of HiddenBench/Werewolf (or give a weaker multi-source concentration bound), and whether the empirical herding floor is consistent with concentration on a small set rather than specifically a pair. Without this, the quantitative Neff≤8 prediction is only loosely connected to the benchmark herding floors.
minor comments (5)
  1. [Figure 1] Figure 1 is dense; the herding vs. wisdom cartoons and the operator bridge formula compete for space. Consider splitting the schematic from the task vignette.
  2. [§4.3, Theorem 4.2] Notation: π is social power and ν is the consensus weight of C; both are stationary-like objects. A one-line reminder at first use of ν_C(β) would help.
  3. [Table 1, §5.3] Table 1 and Appendix 13.2: the equalizer can slightly raise wide-β error in some cells (e.g., k=4). A brief sentence on when equalization can overshoot an already-balanced allocation would prevent misreading.
  4. [Appendix 13.8] Boundary environments (Debate, GovSim, MARBLE) are useful; the negative MARBLE placement result is honest. Consider moving the full boundary table into the main text or a short dedicated subsection so the scope of the theory is clearer.
  5. [Throughout] Minor typos: “APREPRINT” headers; occasional spacing in math (e.g., N_eff formatting). Standard copy-edit pass.

Circularity Check

1 steps flagged · score 2.0 of 10

Mostly non-circular: Neff/accuracy theory is derived from the tempered operator, not fitted to the target; only mild residual-fitting in the bridge diagnostic.

  1. fitted input called prediction [§7 Measuring δ; §5.2 / Fig. 4 / App. 13.3 bridge certification]
    "The anchoring weights (λi) are not assumed known: they are fit per agent by least squares over the run, minimizing ∑t δ̂i(t)² over λi∈[0,1). ... the measured sup-norm proxy↔LLM discrepancy is at or under the theoretical ceiling δ/(1−λ) in 254/254 cells (0 violations) ... Because λi is fit from the T residuals of one run, δT is the worst-case residual after each agent's best-fit anchoring, so "zero violations" is a statement about per-agent best-fit ceilings rather than about a prescribed λ."

    λ is chosen to minimize the residuals that define δ; the reported bridge success (0 violations of δ/(1−λ)) is then evaluated at those fitted λ values. The ceiling is therefore partly optimized by the same data used to check it. This is a mild diagnostic circularity, not a reduction of Theorems 4.2–4.3 or of the external accuracy results.

full rationale

The load-bearing theory (Theorems 4.2–4.3) starts from the explicit realized-influence operator C(β)=Γ_β(W̄⊙π) and derives Neff bounds via tempered-allocation concentration, stationary perturbation, and column-defect control. Those derivations are ordinary consequences of the model definitions, not reductions of the target accuracy into the axioms. Empirical herding is scored by collective accuracy against external task ground truth on HiddenBench/Werewolf/AgentsNet while β is an operator control, not a fit to accuracy. The only mild circularity is diagnostic: anchoring weights λ_i are least-squares fit from the same residuals that enter the bridge ceiling, after which the paper reports 0 violations of the fitted ceiling. That is standard residual fitting and is confined to bridge certification on the anchored testbed; it does not force the Neff theorems or the primary accuracy swings. No self-citation uniqueness chain, no ansatz smuggled via overlapping authors, and no renaming of a known result as the central claim. Score 2 for one non-load-bearing fitted-input diagnostic.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on classical opinion-dynamics structure plus three modeling choices specific to LLMs: finite tempered attention, Katz–Bonacich social power on exposure, and δ-approximate anchored best-response emissions. Free knobs (β, ζ, λ, Sinkhorn steps, coverage threshold) are operator or fit parameters. The invented objects are the SNLA tuple and the priced realized-influence operator; they are operational definitions with measurable handles (C, Neff, accuracy), not unfalsifiable mediators.

free parameters (5)
  • attention width β
    Primary control knob of the realized-influence softmax; swept experimentally (e.g. 0.1–10) rather than derived from first principles.
  • social-power damping ζ
    Katz–Bonacich damping; defaults 0.6–0.7; herding vanishes at ζ=0 by construction of the gate.
  • anchoring weights λ_i
    Required for non-vacuous bridge; main benchmarks use λ=1; bridge certification fits λ per agent by least squares on residuals.
  • Sinkhorn iterations k / online price updates
    Equalizer depth; default k=4; one update per round in the online protocol; column defect is measured not predicted a priori.
  • context coverage threshold 0.9
    Hard inclusion cutoff for peers in the prompt; implements β as an information bottleneck.
assumptions (4)
  • domain assumption Private signals are exogenous and independent of the realized influence sequence {C(t)} (Assumption 7.3).
    Licenses treating q_T as fixed under noise when computing variance; stated as matching fixed-topology environments.
  • domain assumption LLM emissions are δ-approximate anchored best responses evaluated at own beliefs (Assumption 7.2), with λ_max < 1 for the bridge.
    Connects operational LLM beliefs to the linear proxy; fails to cover unanchored DeGroot runs used in main accuracy plots.
  • standard math Classical DeGroot / Friedkin–Johnsen updating and Golub–Jackson wisdom characterization as the λ→1 / β→∞ reference.
    Recovered as special cases when attention is unselective and power uniform.
  • ad hoc to paper Realized influence is row-wise tempered allocation of exposure×power scores (Eq. 1).
    Modeling choice defining SNLA; not forced by LLM internals beyond limited context.
invented entities (2)
  • SNLA realized-influence operator C(β,ζ) independent evidence
    purpose: Maps visible exposure and social power into the weights that actually update beliefs under finite attention.
    Core modeling object; operationally measurable from event logs and prompt construction.
  • Exposure prices y_j (equalized allocator) independent evidence
    purpose: Decentralized column balancing so no source occupies unbounded context share.
    Equivalent to Sinkhorn; falsifiable via measured column defect and accuracy recovery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Social Networks of LLM Agents." pith.science (2026). https://pith.science/paper/QXIJJQFV

@misc{pith2026260703695,
  author       = {Pith},
  title        = {Pith review of: Social Networks of LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXIJJQFV}},
  note         = {Machine review of arXiv:2607.03695}
}
read the original abstract

Large language model (LLM) agents are increasingly deployed in interacting populations, raising the question of what such populations come to believe collectively. Whether a population aggregates genuine knowledge or collapses into a false consensus directly affects how much such systems can be trusted. Classical social-network models assume that the network itself determines how beliefs combine. This assumption breaks down for LLM agents, whose limited attention takes in only part of what they are exposed to, so these models overstate how much information a population actually pools and cannot tell genuine consensus from herding. We introduce SNLA, a framework that models how much each agent actually influences others, rather than merely how the network connects them. This influence depends on each agent's position in the network and on how sharply attention focuses. Theoretically, we show on a tractable proxy that narrow attention causes herding, where the effective sample size stays bounded regardless of population size, while wide attention recovers wisdom-of-crowds behavior only when the exposure graph is undirected and degree-regular. Empirically, a controlled testbed validates these predictions directly, and the herding-wisdom transition reproduces on operator-controlled variants of three multi-agent LLM benchmarks.

Figures

Figures reproduced from arXiv: 2607.03695 by the authors.

Figure 1
Figure 1. From exposure to realized influence. The visible exposure graph W¯ (left) records what each agent could see; edges are exposure weights and node a3 carries high social power π. The realized-influence operator Ci· = Γβ(W¯ i· ⊙π) reweights each reader’s exposure by source social power and concentrates it through a finite attention width β. The graph C (right) depends sharply on β: narrow attention (β=0.3) collapses in… view at source ↗
Figure 2
Figure 2. One influence operator, three recognized environments, two facets. (A, B) Collective accuracy vs. context width β on HiddenBench (n=24) and Werewolf (n=16). Orange is the baseline allocator, blue is the equalizer, and light blue is the control that removes the dominant-wrong source. Dotted lines show individual and oracle reference levels; shaded bands in (A, B) span the seed-to-seed min–max over 5 seeds. (C) Colour… view at source ↗
Figure 3
Figure 3. The equalizer clears herding online, round by round. Per-round collective accuracy at narrow β on HiddenBench (n=24) and Werewolf (n=16). Orange is the baseline; grey dashed is the operator, obtained by clearing the equalized operator to its fixed point each round; blue is online pricing, which applies one price update per discussion round. Shaded bands span the min–max over 5 seeds. (c) Fraction of the narrow-β gap… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The proxy–LLM bridge (Theorem 4.1) on real Qwen data. (a) Measured sup-norm discrepancy supt∥bLLM − bBR∥∞ vs. the theoretical ceiling δT /(1 − λmax), pooling 254 cells — one per (arm, β, seed) with λmax < 1 — across four 8-seed testbed runs spanning Qwen2.5-3B and -7B …
Figure 5
Figure 5. Figure 5: Mechanism ablations (visual). (a) the equalized allocator’s column defect falls monotonically with Sinkhorn iterations k, from 0.14 at k=1 to 0.037 at k=4; (b) herding is gated by social power, absent at ζ=0 and present for any ζ>0 (cf [PITH_FULL_IMAGE:figures/full_fi…
Figure 6
Figure 6. Figure 6: The closed-form two-regime, realized on the LLM testbed. Baseline collective error shows the two regimes across β while the equalized operator holds error low; Neff rises from the narrow-β floor toward the degree ceiling (Theorems 4.2, 4.3). Real Qwen2.5-7B. (flat narr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 12 linked inside Pith

  1. [1]

    Opinion dynamics and learning in social networks.Dynamic Games and Applications, 1(1):3–49, 2011

    Daron Acemoglu and Asuman Ozdaglar. Opinion dynamics and learning in social networks.Dynamic Games and Applications, 1(1):3–49, 2011

    2011

  2. [2]

    Multiagent collaboration attack: Investigating adversarial attacks in large language model collaborations via debate

    Alfonso Amayuelas, Xianjun Yang, Antonis Antoniades, Wenyue Hua, Liangming Pan, and William Yang Wang. Multiagent collaboration attack: Investigating adversarial attacks in large language model collaborations via debate. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 6929–6948, 2024

    2024

  3. [3]

    Learning from neighbours.The review of economic studies, 65(3):595–621, 1998

    Venkatesh Bala and Sanjeev Goyal. Learning from neighbours.The review of economic studies, 65(3):595–621, 1998

    1998

  4. [4]

    A simple model of herd behavior.The quarterly journal of economics, 107(3):797–817, 1992

    Abhijit V Banerjee. A simple model of herd behavior.The quarterly journal of economics, 107(3):797–817, 1992

    1992

  5. [5]

    A theory of fads, fashion, custom, and cultural change as informational cascades.Journal of political Economy, 100(5):992–1026, 1992

    Sushil Bikhchandani, David Hirshleifer, and Ivo Welch. A theory of fads, fashion, custom, and cultural change as informational cascades.Journal of political Economy, 100(5):992–1026, 1992

    1992

  6. [6]

    Power and centrality: A family of measures.American journal of sociology, 92(5):1170–1182, 1987

    Phillip Bonacich. Power and centrality: A family of measures.American journal of sociology, 92(5):1170–1182, 1987

    1987

  7. [7]

    The pagerank citation ranking: bringing order to the web.Proceedings of ASIS, 1998, 98:161–172, 1998

    Sergey Brin. The pagerank citation ranking: bringing order to the web.Proceedings of ASIS, 1998, 98:161–172, 1998

    1998

  8. [8]

    Chateval: Towards better llm-based evaluators through multi-agent debate

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. Chateval: Towards better llm-based evaluators through multi-agent debate. InInternational conference on learning representations, volume 2024, pages 9079–9093, 2024

    2024

Show all 52 references
  1. [9]

    Herd behavior: Investigating peer influence in llm-based multi-agent systems.arXiv preprint arXiv:2505.21588, 2025

    Young-Min Cho, Sharath Chandra Guntuku, and Lyle Ungar. Herd behavior: Investigating peer influence in llm-based multi-agent systems.arXiv preprint arXiv:2505.21588, 2025

    2025 arXiv

  2. [10]

    Simulating opinion dynamics with networks of llm-based agents

    Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka, Siddharth Suresh, Robert Hawkins, Sijia Yang, Dhavan Shah, Junjie Hu, and Timothy Rogers. Simulating opinion dynamics with networks of llm-based agents. InFindings of the association for computational linguistics: NAACL 2024, pag...

    2024

  3. [11]

    Unified scaling laws for routed language models

    Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. InInternational conference on machine learning, pages 4057–4086...

    2022

  4. [12]

    I-divergence geometry of probability distributions and minimization problems.The annals of probability, pages 146–158, 1975

    Imre Csiszár. I-divergence geometry of probability distributions and minimization problems.The annals of probability, pages 146–158, 1975

    1975

  5. [13]

    Language understanding as a constraint on consensus size in llm societies.arXiv preprint arXiv:2409.02822, 2024

    Giordano De Marzo, Claudio Castellano, and David Garcia. Language understanding as a constraint on consensus size in llm societies.arXiv preprint arXiv:2409.02822, 2024

    2024 arXiv

  6. [14]

    Reaching a consensus.Journal of the American Statistical association, 69(345):118–121, 1974

    Morris H DeGroot. Reaching a consensus.Journal of the American Statistical association, 69(345):118–121, 1974

    1974

  7. [15]

    Improving factuality and reasoning in language models through multiagent debate

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

    2024

  8. [16]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    2022

  9. [17]

    On the scaling of multidimensional matrices.Linear Algebra and its applications, 114:717–735, 1989

    Joel Franklin and Jens Lorenz. On the scaling of multidimensional matrices.Linear Algebra and its applications, 114:717–735, 1989

    1989

  10. [18]

    Social influence and opinions.Journal of mathematical sociology, 15 (3-4):193–206, 1990

    Noah E Friedkin and Eugene C Johnsen. Social influence and opinions.Journal of mathematical sociology, 15 (3-4):193–206, 1990

    1990

  11. [19]

    V ox populi, 1907

    Francis Galton. V ox populi, 1907

    1907

  12. [20]

    S3: Social-network simulation system with large language model-empowered agents.arXiv preprint arXiv:2307.14984, 2023

    Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Jinghua Piao, Huandong Wang, Depeng Jin, and Yong Li. S3: Social-network simulation system with large language model-empowered agents.arXiv preprint arXiv:2307.14984, 2023

    2023 arXiv

  13. [21]

    Naive learning in social networks and the wisdom of crowds.American Economic Journal: Microeconomics, 2(1):112–149, 2010

    Benjamin Golub and Matthew O Jackson. Naive learning in social networks and the wisdom of crowds.American Economic Journal: Microeconomics, 2(1):112–149, 2010

    2010

  14. [22]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  15. [23]

    Agentsnet: Coordination and collaborative reasoning in multi-agent llms.arXiv preprint arXiv:2507.08616, 2025

    Florian Grötschla, Luis Müller, Jan Tönshoff, Mikhail Galkin, and Bryan Perozzi. Agentsnet: Coordination and collaborative reasoning in multi-agent llms.arXiv preprint arXiv:2507.08616, 2025

    2025 arXiv

  16. [24]

    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 2024.arXiv preprint arXiv:2402.01680, 10, 2024

    2024 arXiv

  17. [25]

    Metagpt: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Steven Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, volume 2024...

    2024

  18. [26]

    Survey sampling

    Leslie Kish. Survey sampling. 1965

    1965

  19. [27]

    American Mathematical Society, 2026

    David A Levin and Yuval Peres.Markov chains and mixing times. American Mathematical Society, 2026

    2026

  20. [28]

    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

    2023

  21. [29]

    Hiddenbench: Assessing collective reasoning in multi-agent llms via hidden profile tasks.arXiv preprint arXiv:2505.11556, 2025

    Yuxuan Li, Aoi Naito, and Hirokazu Shirado. Hiddenbench: Assessing collective reasoning in multi-agent llms via hidden profile tasks.arXiv preprint arXiv:2505.11556, 2025

    2025 arXiv

  22. [30]

    Encouraging divergent thinking in large language models through multi-agent debate

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. InProceedings of the 2024 conference on empirical methods in natural language proc...

    2024

  23. [31]

    The role of the group generalized inverse in the theory of finite markov chains.Siam Review, 17 (3):443–464, 1975

    Carl D Meyer, Jr. The role of the group generalized inverse in the theory of finite markov chains.Siam Review, 17 (3):443–464, 1975

    1975

  24. [32]

    Network formation and dynamics among multi-llms.PNAS nexus, 4(12): pgaf317, 2025

    Marios Papachristou and Yuan Yuan. Network formation and dynamics among multi-llms.PNAS nexus, 4(12): pgaf317, 2025

    2025

  25. [33]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology, pages 1–22, 2023

    2023

  26. [34]

    Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society

    Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al. Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society. 2025

    2025

  27. [35]

    Cooperate or collapse: Emergence of sustainable cooperation in a society of llm agents.Advances in Neural Information Processing Systems, 37:111715–111759, 2024

    Giorgio Piatti, Zhijing Jin, Max Kleiman-Weiner, Bernhard Schölkopf, Mrinmaya Sachan, and Rada Mihalcea. Cooperate or collapse: Emergence of sustainable cooperation in a society of llm agents.Advances in Neural Information Processing Systems, 37:111715–111759, 2024

    2024

  28. [36]

    Consensagent: Towards efficient and effective consensus in multi-agent llm interactions through sycophancy mitigation

    Priya Pitre, Naren Ramakrishnan, and Xuan Wang. Consensagent: Towards efficient and effective consensus in multi-agent llm interactions through sycophancy mitigation. InFindings of the Association for Computational Linguistics: ACL 2025, pages 22112–22133, 2025

    2025

  29. [37]

    A tutorial on modeling and analysis of dynamic social networks

    Anton V Proskurnikov and Roberto Tempo. A tutorial on modeling and analysis of dynamic social networks. part i.Annual Reviews in Control, 43:65–79, 2017

    2017

  30. [38]

    Chatdev: Communicative agents for software development

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 1:...

    2024

  31. [39]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  32. [40]

    Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  33. [41]

    Diagonal equivalence to matrices with prescribed row and column sums.The American Mathematical Monthly, 74(4):402–405, 1967

    Richard Sinkhorn. Diagonal equivalence to matrices with prescribed row and column sums.The American Mathematical Monthly, 74(4):402–405, 1967

    1967

  34. [42]

    Systematic biases in llm simulations of debates

    Amir Taubenfeld, Yaniv Dover, Roi Reichart, and Ariel Goldstein. Systematic biases in llm simulations of debates. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 251–267, 2024

    2024

  35. [43]

    Auxiliary-loss-free load balancing strategy for mixture-of-experts.arXiv preprint arXiv:2408.15664, 2024

    Lean Wang, Huazuo Gao, Chenggang Zhao, Xu Sun, and Damai Dai. Auxiliary-loss-free load balancing strategy for mixture-of-experts.arXiv preprint arXiv:2408.15664, 2024

    2024 arXiv

  36. [44]

    Autogen: Enabling next-gen llm applications via multi-agent conversation.arXiv preprint arXiv:2308.08155, 2023

    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 conversation.arXiv preprint arXiv:2308.08155, 2023

    2023 arXiv

  37. [45]

    Language agents with reinforcement learning for strategic play in the werewolf game.arXiv preprint arXiv:2310.18940, 2023

    Zelai Xu, Chao Yu, Fei Fang, Yu Wang, and Yi Wu. Language agents with reinforcement learning for strategic play in the werewolf game.arXiv preprint arXiv:2310.18940, 2023

    2023 arXiv

  38. [46]

    Multiagentbench: Evaluating the collaboration and competition of llm agents

    Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Daisy Zhe Wang, Zhenhailong Wang, Cheng Qian, Robert Tang, Heng Ji, et al. Multiagentbench: Evaluating the collaboration and competition of llm agents. InProceedings of the 63rd Annual Meeting of the Association ...

    2025

  39. [47]

    (Wide, wisdom of crowds.)If B=B ⊤ is irreducible, then C(β)→D −1B and νC(β) →d/(1 ⊤d) as β→ ∞, so: lim β→∞ Neff(β) = (P i di)2 P i d2 i , which equals n if and only ifB is degree-regular; for finiteβ the consensus weight obeys ∥νC(β) ∥2 2 −∥ν ⋆∥2 2 ≤ 2κB(eL/β −1) with ν⋆ =d/(1...

  40. [48]

    Then with ρ(β) = (d max −1)e −g/β, wheneverρ(β)<1, Neff(β)≤ 2 (1−ρ(β)) 2 , soN eff(β)≤8wheneverρ(β)≤ 1 2, i.e.β≤g /log(2(d max −1)), independently ofn

    (Narrow, herding.)Assume B is irreducible and the scores admit a unique dominant pair: j⋆ i = arg maxk sik is unique for every row, with j⋆ i =h for all i̸=h and j⋆ h =g̸=h . Then with ρ(β) = (d max −1)e −g/β, wheneverρ(β)<1, Neff(β)≤ 2 (1−ρ(β)) 2 , soN eff(β)≤8wheneverρ(β)≤ 1...

    2026

  41. [49]

    We bound ∥ν∥2 2 from above. For νh, using 1 1−x ≤1 + x 1−x with x=ρ/2, νh ≤ 1 2−ρ = 1 2 · 1 1−ρ/2 ≤ 1 2 1 + ρ/2 1−ρ/2 ≤ 1 2 1 + 4 7 ρ ≤ 1 2 + 1 3 ρ, where ρ/2 1−ρ/2 ≤ ρ/2 7/8 = 4 7 ρ at ρ≤ 1 4, and 1 2 · 4 7 = 2 7 ≤ 1

  42. [50]

    [weight 0.31] Agent 7

    By equation 21, νg ≤ν h +ρ≤ 1 2 + 4 3 ρ. The mass on Sc contributes at mostP k∈S c ν2 k ≤(P k∈S c νk)2 ≤ρ 2. Summing, ∥ν∥2 2 ≤ 1 2 + 1 3 ρ 2 + 1 2 + 4 3 ρ 2 +ρ 2 = 1 2 + ρ 3 + 4ρ 3 + ρ2 9 + 16ρ2 9 +ρ 2 = 1 2 + 5 3 ρ+ 26 9 ρ2. Atρ≤ 1 4, 26 9 ρ2 ≤ 26 9 · 1 4 ρ= 26 36 ρ≤ 5 6 ρ, s...

    2026

  43. [51]

    Safe Haven After the Spill

    {option 2} ... First, in 1-2 sentences, state the specific fact(s) YOU were given that others may not have. Then weigh all available facts and choose. End with exactly one line: BELIEF: <integer> where <integer> is the option number of your current best decision (digits only)....

    2026

  44. [52]

    Maple Lodge is the listed safe site

    Cedar Station First, in 1-2 sentences, state the specific fact(s) YOU were given that others may not have. Then weigh all available facts and choose. End with exactly one line: BELIEF: <integer> where <integer> is the option number of your current best decision (digits only). ...

    2026

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.