Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Modern multi-agent LLM systems fail to explore their peers, locking onto early favorites and raising regret; explicit guided peer selection fixes this, and the payoff grows with how different the agents are.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 06:00 UTC pith:VTZ3WKHP

load-bearing objection Solid empirical diagnosis that LLMs polarize on peer selection, plus a practical LinUCB fix; theory is standard bandit analysis on a surrogate, not a tight POSG solution. the 3 major comments →

arxiv 2607.11250 v1 pith:VTZ3WKHP submitted 2026-07-13 cs.MA cs.AI

Multi-Agent LLMs Fail to Explore Each Other

Hyeong Kyu Choi , Jiatong Li , Wendi Li , Xin Eric Wang , Sharon Li This is my paper
classification cs.MA cs.AI
keywords multi-agent explorationLLM agentscontextual banditspeer selectionLinUCBcapability diversitypartially observable stochastic gameregret
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that exploration is a missing foundation of reliable multi-agent autonomy for large language models. When agents must choose whom to query under uncertainty about peer capabilities or private context, current LLMs do not balance trial and commitment: they polarize early, stick with the first plausible peer, and often never correct a bad choice. The authors cast this as Multi-Agent Exploration—a partially observable multi-agent game in which each agent must probe others to learn who is useful—and show the failure even in a simple two-peer bandit and again in larger systems with split context or mixed model strengths. They propose Multi-Agent Contextual Exploration (MACE), a light LinUCB-style peer selector that scores candidates with relational features (answer diversity, peer distinctiveness, past success, interaction stage) plus an uncertainty bonus. Empirically, MACE spreads selections more evenly, cuts cumulative regret relative to random and to pure prompting, improves answer quality across HotpotQA-style and math/science tasks, and transfers frozen parameters to a harder multi-hop set. The theory states that a non-exploring greedy policy pays linear regret proportional to capability diversity, while MACE keeps sublinear regret, so exploration’s value rises as agents specialize.

Core claim

Current LLM agents lack an innate capacity for multi-agent exploration: under uncertainty about peers they commit prematurely and polarizedly, which produces suboptimal coordination and higher regret. Explicit algorithmic structure—not more prompting—is needed to restore exploration, and the resulting benefit scales with how diverse the agent pool is.

What carries the argument

Multi-Agent Contextual Exploration (MACE): each agent independently treats peer choice as a linear contextual bandit whose features encode relational utility of a query (n-gram response diversity, peer distinctiveness in the group, historical reward, normalized round, bias), then selects via LinUCB’s optimistic score so under-tested peers in the current relational context get tried before exploitation.

Load-bearing premise

That the hard multi-agent game can be safely reduced to independent linear bandits whose hand-built relational features capture peer non-stationarity well enough for standard optimism-based regret bounds to hold.

What would settle it

On the same HotpotQA distractor and mixed-model GPQA/Math500 protocols, show that pure in-context exploration (or any prompt-only policy with the same history) matches or beats MACE in cumulative regret offset and final-round accuracy when capability diversity is high; or show that MACE’s regret does not stay near the claimed O(sqrt(T log T)) envelope while a greedy no-bonus policy stays near linear in diversity.

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

If this is right

  • Prompt-only multi-agent setups will keep locking onto early peers and can underperform even random routing when contexts or skills are split.
  • Reliable open multi-agent systems need an explicit exploration layer for peer selection, not only role prompts or fixed debate graphs.
  • As agent pools become more specialized, the performance gap between exploring and non-exploring policies will widen roughly linearly in diversity.
  • Interaction policies learned under MACE can transfer frozen to related tasks, so exploration can be amortized rather than restarted every benchmark.
  • Even strong frontier models in a mixed pool still gain from guided peer discovery, so the failure is structural rather than a small-model artifact.

Where Pith is reading between the lines

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

  • Fixed communication topologies in multi-agent debate and role-play systems may be silently inheriting the same premature-commitment pathology unless peer choice is reopened with exploration incentives.
  • If relational features are the real lever, richer online signals (tool outcomes, trust calibration, delayed team reward) could extend MACE from answer quality to long-horizon coordination without solving the full multi-agent game.
  • Homogeneous strong-model pools will show small regret gains from exploration but may still show accuracy gains via better information mixing—suggesting separate metrics for selection efficiency versus collective reasoning quality.
  • Large-N deployments (hundreds of agents) will likely need hierarchical or clustered variants of MACE, because flat per-peer LinUCB scales poorly when useful specialists are rare.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper argues that modern LLM agents fail to explore peers in multi-agent settings, producing premature, polarized peer selection and elevated regret. It formalizes Multi-Agent Exploration as a POSG in which agents must probe heterogeneous peers under partial observability, then proposes MACE: independent per-agent LinUCB peer selection over a hand-designed 9-dimensional relational feature vector (n-gram response diversity, peer distinctiveness, historical reward, normalized round, bias). Empirically, MACE improves peer-selection coverage, cumulative regret, and task metrics (HotpotQA EM/F1; Math500/GPQA accuracy) relative to in-context exploration, random, and fixed-topology baselines under both contextual and parametric diversity, with transfer of frozen parameters to 2WikiMultihopQA and supporting homogeneous/stronger-model ablations. Theoretically, standard LinUCB analysis yields Õ(√T log T) regret for MACE versus Ω(δT) for a greedy non-exploring policy, so the exploration benefit grows with capability diversity δ.

Significance. The core empirical finding—that even strong frontier models lock onto peers early and that prompting alone does not fix this—is timely and practically important for multi-agent LLM systems. The multi-setting evaluation (contextual HotpotQA, parametric Math500/GPQA, transfer, homogeneous and stronger-model ablations), clear baselines, regret-offset plots, and planned code release are real strengths. The formalization of multi-agent exploration and the diversity–exploration link (Corollary 1) are useful organizing ideas. The theory is standard self-normalized linear-bandit analysis rather than a new multi-agent result; its value depends on how faithfully the POSG is reduced to independent linear bandits. If the empirical claims hold under more realistic feedback, the paper would be a solid contribution to multi-agent LLM reliability.

major comments (3)
  1. [Section 4; Theorems 1–2; Appendix E] Section 4 and Appendix E: The POSG is reduced to independent linear contextual bandits (Eqs. 2–5) with a fixed 9-d feature map, and Theorems 1–2 import stationary LinUCB guarantees under Assumptions 1–3. Peers are adaptive agents whose responses evolve with their own histories, so μ_{i,a,t} need not remain linear in a fixed map, and the filtration/noise conditions of Abbasi-Yadkori et al. (2011) may fail. The O(√T log T) bound and the Ω(δT) exploration benefit are therefore guarantees for the surrogate bandit, not for the original POSG. The manuscript should state this scope explicitly and either weaken the theoretical claim about the multi-agent problem or provide evidence that the feature map keeps the linear residual well-behaved under peer adaptivity.
  2. [Appendix B.2; Section 5] Appendix B.2, Eq. (20): The reward used to update LinUCB is a blend of score improvement and absolute post-interaction score against ground truth (token F1 or exact match). Peer-selection learning therefore depends on oracle correctness feedback at every round. This is load-bearing for MACE’s reported gains and for the autonomy narrative in the abstract and §8. The paper should either (i) evaluate a label-free reward (e.g., self-consistency, peer agreement, or delayed batch labels) or (ii) clearly restrict claims to settings with verifiable intermediate rewards and discuss how the method would be deployed without them.
  3. [Appendix B.1; Appendix D.6; Figure 3–4] Appendix B.1 / D.6: The 9 relational features are hand-designed and free parameters of the method (along with α and λ). Feature-weight plots are informative but not ablations. Without leave-one-feature-out or simpler baselines (e.g., scalar UCB on historical reward alone; diversity-only), it is hard to know whether the relational design is necessary for the gains in Figures 3–4 or whether a much simpler exploration bonus would suffice. A short ablation table would substantially strengthen the method claim.
minor comments (5)
  1. [Section 2; Appendix A] Figure 1 and Appendix A: The two-armed bandit motivating experiment is clear, but the main text should note temperature (1.2 for Qwen) and that peers answer with fixed Bernoulli success rates, so this is pure selection under stationary arms—not yet multi-agent adaptivity.
  2. [Figure 4] Figure 4 caption and regret definition: “cumulative regret offset (Baseline−Random)” is easy to misread as higher-is-better; the text correctly says lower is better, but a single consistent sign convention in all panels would help.
  3. [Appendix C.4; Section 5.2] In-Context Exploration prompt (Appendix C.4) dumps the full 9-d feature vector into the LLM. A control with only counts and mean rewards (no engineered features) would better isolate “prompting fails” from “prompting with MACE’s features fails.”
  4. [Section 5; Appendix F] N and R are small (N=10/R=5 contextual; N=4/R=3 parametric). Appendix F acknowledges large-N limits; a brief scaling note or cost discussion in the main text would set expectations.
  5. [Abstract; Eq. (1); Algorithm 1] Typos/notation: “Multi- Agent” spacing in the abstract; Regret definition (1) uses μ^{(t)}_{i,a} without defining time-dependence of the mean before the bandit reduction; Algorithm 1 uses 9-d features while the main text writes R^d generically.

Circularity Check

0 steps flagged

No significant circularity: empirical claims rest on external benchmarks and baselines; theoretical regret bounds are standard LinUCB arguments under stated assumptions, not tautologies or self-fitted predictions.

full rationale

The paper's central claims (LLM agents exhibit polarized premature commitment; MACE improves exploration and task metrics; exploration value scales with capability diversity δ) are not true by construction. Empirical results compare peer-selection histograms, cumulative regret offsets, and round-wise EM/F1/accuracy against Random, Pre-defined, and In-Context Exploration baselines on HotpotQA, Math500, GPQA, and a transfer set (2WikiMultiHopQA); these are external measurements, not rearrangements of fitted inputs. MACE is an explicit LinUCB reduction (Eqs. 2–5) with hand-designed 9-d relational features and a blended reward; feature weights and α are learned or swept (Appendix D.2, D.6), not reverse-engineered to force the conclusion. Theorems 1–2 and Corollary 1 import classical linear-bandit concentration (Abbasi-Yadkori et al.) under Assumptions 1–3; the O(√T log T) upper bound and Ω(δT) lower bound for the non-exploring policy follow from those assumptions rather than from defining the target in terms of itself. Capability diversity δ is defined (Def. 1) and then assumed to lower-bound expected gaps (Ass. 3); this is a modeling hypothesis, not a circular derivation. No load-bearing uniqueness theorem is imported from overlapping authors, no ansatz is smuggled via self-citation, and no fitted parameter is renamed a “prediction.” The POSG-to-independent-bandit reduction is an approximation whose fidelity can be questioned on correctness grounds, but that is not circularity under the stated criteria. The derivation chain is therefore self-contained against external evidence.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard linear-bandit concentration, a modeling reduction from POSG to independent contextual bandits, hand-designed relational features, a blended improvement+absolute reward, and the capability-diversity gap assumption. Free parameters are the usual bandit knobs (α, λ) plus feature engineering choices; no new physical entities are postulated.

free parameters (4)
  • exploration coefficient α = typically 0.5–1.0 (sensitivity in App. D.2)
    Controls the UCB bonus; sensitivity analysis shows intermediate values (0.5–1.0) work best; chosen by hand/sweep rather than derived.
  • ridge regularization λ = 1.0
    Fixed to 1.0 across settings; standard but free.
  • 9-dimensional relational feature design
    n-gram diversities, distinctiveness, historical reward, normalized round, bias are hand-crafted; weights are learned but the feature set itself is a free modeling choice.
  • blended reward weights (1/2 improvement + 1/2 absolute score) = 1/2, 1/2
    Reward definition mixes score improvement and absolute post-interaction score with equal weights; design choice affecting the bandit target.
axioms (4)
  • standard math Bounded features/parameters and R-sub-Gaussian reward noise (Assumptions 1–2)
    Standard self-normalized martingale conditions for LinUCB regret (Abbasi-Yadkori et al.).
  • domain assumption Capability diversity δ lower-bounds expected suboptimal-peer penalty by βδ (Assumption 3)
    Links latent capability vectors to reward gaps; needed for the linear lower bound on non-exploring regret.
  • ad hoc to paper Independent per-agent linear contextual bandit reduction adequately approximates the POSG with adaptive peers
    Core modeling step in Section 4; non-stationarity is only encoded via the relational features, not jointly modeled.
  • domain assumption Ground-truth task scores are available to compute the bandit reward during the trial-and-error phase
    Reward uses EM/F1/accuracy against gold answers; realistic only when labels exist for training the selector.
invented entities (2)
  • Multi-Agent Exploration problem (POSG formalization) no independent evidence
    purpose: Names and formalizes the peer-probing decision problem for heterogeneous LLM agents.
    New problem statement for this setting; not an ontological entity beyond a modeling frame.
  • MACE (Multi-Agent Contextual Exploration) with relational features no independent evidence
    purpose: Tractable peer-selection policy that injects explicit exploration.
    Algorithmic construct; evaluated empirically and theoretically inside the paper.

pith-pipeline@v1.1.0-grok45 · 51912 in / 3175 out tokens · 40076 ms · 2026-07-14T06:00:09.000328+00:00 · methodology

0 comments
read the original abstract

Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents fail to do so, often exhibiting myopic and polarized interaction patterns that lead to suboptimal coordination and increased regret. We formalize this challenge as the Multi-Agent Exploration problem, modeling it as a partially observable stochastic game (POSG) problem in which agents must probe peers to infer their capabilities and identify effective interaction strategies. To address this, we introduce Multi- Agent Contextual Exploration (MACE), a lightweight framework that explicitly promotes exploration through structured peer selection. Across both contextual and parametric diversity settings, MACE substantially improves exploration behavior and downstream task performance. We further show theoretically that the value of exploration increases with agent diversity. Overall, our results highlight a fundamental limitation of current LLM agents and underscore the importance of explicitly guided exploration for reliable multi-agent autonomy. Code will be released in https://github.com/deeplearning-wisc/mace

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 19 linked inside Pith

  1. [1]

    Achiam, S

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

    Pith/arXiv arXiv

  2. [2]

    Agashe, Y

    S. Agashe, Y . Fan, A. Reyna, and X. E. Wang. Llm-coordination: evaluating and analyzing multi-agent coor- dination abilities in large language models. InFindings of the Association for Computational Linguistics: NAACL 2025, pages 8038–8057,

    2025

  3. [3]

    Amayuelas, J

    A. Amayuelas, J. Yang, S. Agashe, A. Nagarajan, A. Antoniades, X. E. Wang, and W. Wang. Self-resource allocation in multi-agent llm systems.arXiv preprint arXiv:2504.02051,

    Pith/arXiv arXiv

  4. [4]

    Becker, L

    J. Becker, L. B. Kaesberg, N. Bauer, J. P. Wahle, T. Ruas, and B. Gipp. Mallm: Multi-agent large language models framework. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 418–439,

    2025

  5. [5]

    M. Buyl, Y . Fettach, G. Bied, and T. De Bie. Building and measuring trust between large language models. arXiv preprint arXiv:2508.15858,

    Pith/arXiv arXiv

  6. [6]

    M. A. Ferrag, N. Tihanyi, and M. Debbah. From llm reasoning to autonomous ai agents: A comprehensive review.arXiv preprint arXiv:2504.19678,

    Pith/arXiv arXiv

  7. [7]

    D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y . Xie, D. Chen, et al. Agentscope: A flexible yet robust multi-agent platform.arXiv preprint arXiv:2402.14034,

    Pith/arXiv arXiv

  8. [8]

    Grattafiori, A

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

    Pith/arXiv arXiv

  9. [9]

    T. Guo, X. Chen, Y . Wang, R. Chang, S. Pei, N. V . Chawla, O. Wiest, and X. Zhang. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680,

    Pith/arXiv arXiv

  10. [10]

    T. Händler. Balancing autonomy and alignment: A multi-dimensional taxonomy for autonomous llm-powered multi-agent architectures.arXiv preprint arXiv:2310.03659,

    Pith/arXiv arXiv

  11. [11]

    doi: 10.48550/arXiv.2310.03659. X. Ho, A.-K. Duong Nguyen, S. Sugawara, and A. Aizawa. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. InProceedings of the 28th International Conference on Computational Linguistics, pages 6609–6625,

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.03659

  12. [12]

    Z. Hou, J. Tang, and Y . Wang. Halo: Hierarchical autonomous logic-oriented orchestration for multi-agent llm systems.arXiv preprint arXiv:2505.13516,

    Pith/arXiv arXiv

  13. [13]

    Huang, X

    J. Huang, X. Chen, S. Mishra, H. S. Zheng, A. W. Yu, X. Song, and D. Zhou. Large language models cannot self-correct reasoning yet.arXiv preprint arXiv:2310.01798,

    Pith/arXiv arXiv

  14. [14]

    15 Multi-Agent LLMs Fail to Explore Each Other A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. d. l. Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, et al. Mistral 7b.arXiv preprint arXiv:2310.06825,

    Pith/arXiv arXiv

  15. [15]

    humans welcome to observe

    Y . Jiang, Y . Zhang, X. Shen, M. Backes, and Y . Zhang. " humans welcome to observe": A first look at the agent social network moltbook.arXiv preprint arXiv:2602.10127,

    arXiv

  16. [16]

    L. B. Kaesberg, J. Becker, J. P. Wahle, T. Ruas, and B. Gipp. V oting or consensus? decision-making in multi-agent debate. InFindings of the Association for Computational Linguistics: ACL 2025, pages 11640–11671,

    2025

  17. [17]

    X. Li, S. Wang, S. Zeng, Y . Wu, and Y . Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024a. Y . Li, Y . Du, J. Zhang, L. Hou, P. Grabowski, Y . Li, and E. Ie. Improving multi-agent debate with sparse communication topology. InFindings of the Association for Computational Linguistics: EMNL...

    2024

  18. [18]

    Lightman, V

    H. Lightman, V . Kosaraju, Y . Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe. Let’s verify step by step.arXiv preprint arXiv:2305.20050,

    Pith/arXiv arXiv

  19. [19]

    T. Liu, X. Wang, W. Huang, W. Xu, Y . Zeng, L. Jiang, H. Yang, and J. Li. Groupdebate: Enhancing the efficiency of multi-agent debate using group discussion.arXiv preprint arXiv:2409.14051, 2024a. W. Liu, C. Wang, Y . Wang, Z. Xie, R. Qiu, Y . Dang, Z. Du, W. Chen, C. Yang, and C. Qian. Autonomous agents for collaborative task under information asymmetry....

    arXiv

  20. [21]

    org/abs/2604.13151

    URL https://arxiv. org/abs/2604.13151. 16 Multi-Agent LLMs Fail to Explore Each Other J. S. Park, J. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. 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,

    Pith/arXiv arXiv

  21. [22]

    J. Piao, Y . Yan, J. Zhang, N. Li, J. Yan, X. Lan, Z. Lu, Z. Zheng, J. Y . Wang, D. Zhou, et al. Agentsociety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society.arXiv preprint arXiv:2502.08691,

    Pith/arXiv arXiv

  22. [23]

    doi: 10.5281/zenodo.14179129. A. Smit, P. Duckworth, N. Grinsztajn, T. D. Barrett, and A. Pretorius. Should we be going mad? a look at multi-agent debate strategies for llms.arXiv preprint arXiv:2311.17371,

    Pith/arXiv arXiv doi:10.5281/zenodo.14179129

  23. [24]

    Tomašev, M

    N. Tomašev, M. Franklin, and S. Osindero. Intelligent ai delegation.arXiv preprint arXiv:2602.11865,

    arXiv

  24. [25]

    K.-T. Tran, D. Dao, M.-D. Nguyen, Q.-V . Pham, B. O’Sullivan, and H. D. Nguyen. Multi-agent collaboration mechanisms: A survey of llms.arXiv preprint arXiv:2501.06322,

    Pith/arXiv arXiv

  25. [26]

    Q. Wang, T. Wang, Z. Tang, Q. Li, N. Chen, J. Liang, and B. He. Megaagent: A large-scale autonomous llm-based multi-agent system without predefined sops. InFindings of the Association for Computational Linguistics: ACL 2025, pages 4998–5036,

    2025

  26. [27]

    Z. Weng, A. Antoniades, D. Nathani, Z. Zhang, X. Pu, and X. E. Wang. Group-evolving agents: Open-ended self-improvement via experience sharing.arXiv preprint arXiv:2602.04837,

    arXiv

  27. [28]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, et al. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115,

    Pith/arXiv arXiv

  28. [29]

    Z. Yang, P. Qi, S. Zhang, Y . Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2369–2380,

    2018

  29. [30]

    Zhang, K

    Y . Zhang, K. Mei, M. Liu, J. Wang, D. N. Metaxas, X. Wang, J. Hamm, and Y . Ge. Agents in the wild: Safety, society, and the illusion of sociality on moltbook.arXiv preprint arXiv:2602.13284,

    arXiv

  30. [31]

    Zhang, D

    17 Multi-Agent LLMs Fail to Explore Each Other Z. Zhang, D. Wang, N. Chen, R. Mansur, and V . Sarhangian. Comparing exploration-exploitation strategies of llms and humans: Insights from standard multi-armed bandit experiments.arXiv preprint arXiv:2505.09901,

    Pith/arXiv arXiv

  31. [32]

    What is 10 + 32?

    18 Multi-Agent LLMs Fail to Explore Each Other Appendix Table of Contents A Appendix for the Delegation Experiment 20 A.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 A.2 Histograms for All𝑝 A Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 B Appendix for Multi-Agent Contextu...

    2048

  32. [33]

    (16) This feature signals the policy to exploit peers based on historical performance. 23 Multi-Agent LLMs Fail to Explore Each Other Interaction round.We include the normalized per-sample interaction round 𝜏= 𝑟 𝑅 ,(17) where 𝑟∈[0, 𝑅] is the current round and 𝑅 is the total number of interaction rounds for each sample. This feature allows the policy to co...

    2018

  33. [34]

    • Table 1 – 4: Contextual diversity experiment results on HotpotQA for MACE, In-Context Exploration, Pre-defined, and Random

    The tables also include the performance of each individual 30 Multi-Agent LLMs Fail to Explore Each Other agent. • Table 1 – 4: Contextual diversity experiment results on HotpotQA for MACE, In-Context Exploration, Pre-defined, and Random. • Table 9 – 16: Parametric diversity experiment results on GPQA and Math500, for MACE, In-Context Exploration, Pre-def...

    2033

  34. [35]

    Capability diversity and reward gaps.Recall that each peer 𝑎∈ 𝒩 is associated with a latent capability vectorc 𝑎 ∈R 𝑘, and the capability diversity of the agent pool is 𝛿:= 1 𝑁 𝑁∑︁ 𝑎=1 ‖c𝑎 − ¯c‖2, ¯c:= 1 𝑁 𝑁∑︁ 𝑎=1 c𝑎.(27) We assume that capability diversity lower-bounds the expected penalty of selecting a suboptimal peer. Specifically, for some constant𝛽 ...

    2011