Pith. sign in

REVIEW 4 major objections 4 minor 3 cited by

Byzantine-Robust Decentralized Coordination of LLM Agents

T0 review · 4 major / 4 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 DecentLLMs, a leaderless Byzantine-robust protocol, selects higher-quality answers faster than leader-driven quorum voting.

desk verdict A defensible leaderless protocol idea whose central accuracy claim outruns the evidence: the Byzantine-robust aggregation is sound, but the paper needs an evaluator-reliability assumption it never states and its own data contradict. read the letter →

arxiv 2507.14928 v1 pith:QWIVJKYO submitted 2025-07-20 cs.DC cs.AI

Yongrae Jo , Chanik Park This is my paper
classification cs.DCcs.AI
keywords DecentLLMsByzantinefaulttolerancemulti-agentLLMgeometricmedianleaderlessconsensusdecentralizedcoordinationMMLU-Pro
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

DecentLLMs is a leaderless protocol for coordinating multiple LLM agents under Byzantine faults. Worker agents generate answers in parallel; evaluator agents independently score each answer on five quality criteria; the per-answer score vectors are aggregated with the geometric median; and the answer with the highest robust score is returned and recorded on chain. The paper argues this avoids the two main defects of leader-driven schemes: consensus no longer stalls when consecutive leaders are Byzantine, and the final answer is selected from all proposals rather than from whichever proposal a leader happened to make. Experiments on 100 MMLU-Pro problems report 71 correct answers for DecentLLMs versus 64 for two-thirds-quorum voting and 50 for majority quorum, and show roughly constant consensus latency as Byzantine agents are added.

What carries the argument

The carrying object is the geometric median (GM) aggregation of evaluator score vectors. Each evaluator returns a five-dimensional vector of quality scores per worker answer; GM picks the point minimizing summed Euclidean distance to all evaluators' vectors, tolerating up to $f \le \lfloor (n-1)/2 \rfloor$ Byzantine evaluators, compared with $\lfloor (n-2)/2 \rfloor$ for Krum and $\lfloor (n-3)/4 \rfloor$ for Bulyan. This single-round all-to-all evaluation replaces the leader's proposal and quorum vote, and Weiszfeld's algorithm computes it in the implementation.

What would settle it

Run DecentLLMs with only honest workers and honest evaluators on a labeled benchmark and check whether the top geometric-median answer is the gold-correct answer; Table I is already one such check, where incorrect w8 scores 83.0 above correct w3's 55.6, so any run reproducing that pattern with a wrong top pick would refute the claim.

Watch

Extended reading notes

Core claim

The central claim is that abandoning the leader round is enough to make Byzantine-robust multi-agent consensus both faster and higher quality. Concretely, the paper reports that DecentLLMs keeps consensus latency near 221 seconds regardless of the number of Byzantine agents, while rotating-leader and fixed-leader baselines grow roughly linearly, and that geometric-median aggregation continues to select an honest worker's answer until the number of Byzantine evaluators reaches the algorithm's tolerance threshold. The paper concludes that decentralized evaluation with geometric-median aggregation enables faster consensus and ensures that the highest-quality answer is selected even in the presence of Byzantine agents.

Load-bearing premise

The protocol's quality guarantee rests on evaluator scores being faithful signals of correctness, so the highest geometric-median score picks the best answer; Table I already shows a wrong answer outscoring a correct one, so if that fidelity fails the aggregation still returns a poor answer.

Editorial extensions

If this is right

  • If correct, leaderless evaluation removes the view-change bottleneck: consensus stays single-round even when a large share of agents misbehave.
  • Because all worker answers are scored rather than only a leader's proposal, the final answer can be the best available proposal instead of the first one that gathers a quorum.
  • The protocol inherits the geometric median's Byzantine resilience, so an honest-majority evaluator pool keeps selection independent of colluding evaluators up to the stated threshold.
  • The accuracy gain over quorum voting shown on MMLU-Pro suggests leaderless scoring is a usable coordination primitive for open, decentralized LLM deployments.

Reading between the lines

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

  • A testable corollary the paper leaves implicit: the protocol's ceiling is evaluator scoring fidelity, so swapping in weaker evaluators should degrade selection accuracy even with zero Byzantine agents.
  • The same architecture could be applied to code generation or formal reasoning where answers are automatically verifiable; then evaluator reliability could be audited directly, closing the gap the paper acknowledges in Table I.
  • Adding multi-round debate among evaluators to compress honest score variance, which the paper lists as future work, is a natural extension that could be benchmarked against the single-round version.
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

4 major / 4 minor

Summary. The paper proposes DecentLLMs, a leaderless Byzantine-robust coordination protocol for multi-agent LLM systems. Worker agents generate answers in parallel, evaluator agents score each answer on five criteria, and the protocol aggregates the evaluators' score vectors with the geometric median (GM) to select the answer with the highest robust score. The authors claim that this design tolerates Byzantine workers and evaluators, reduces consensus latency compared to leader-driven schemes, and ensures that the highest-quality answer is selected. The experimental evaluation on 100 MMLU-Pro problems reports 71/100 accuracy for DecentLLMs versus 64 for a two-thirds quorum baseline and 50 for a majority-quorum baseline, a roughly constant consensus latency as the number of Byzantine agents grows, and GM-based resilience up to the standard <n/2 threshold.

Significance. If the central claims held, DecentLLMs would be a useful application of Byzantine-robust aggregation to decentralized LLM coordination, addressing the leader-targeting vulnerability of existing leader-driven systems. The paper's strengths are its clear architectural contrast with leader-based protocols, the principled use of the geometric median as a fixed robust estimator (no fitted parameters), and the candid reporting in Sections V-B-3 and V-B-4 that evaluator scores can mis-rank correct and incorrect answers. However, the paper does not introduce new theoretical machinery, and the empirical evidence is weakened by a stylized baseline and a non-end-to-end latency measurement. The significance is therefore moderate: an application-level contribution whose main value would be realized only if the evaluator-reliability problem is explicitly acknowledged and addressed.

major comments (4)
  1. [Section VII and Section III-C-3] The central claim that DecentLLMs 'ensures that the highest-quality answer is selected, even in the presence of Byzantine agents' is not supported by the Byzantine-robustness of the geometric median. Equation (1) robustly aggregates the evaluator-supplied score vectors, but the GM cannot repair a systematic failure of honest evaluator scores to rank answers by true quality. No assumption is stated that honest evaluator scores are monotone in answer quality, and the paper's own evidence contradicts such an assumption: in Table I, the incorrect answer w8 receives a robust score of 83.0 while the correct answer w3 receives 55.6, and Section V-B-3 explicitly admits that 'the evaluators failed to reliably score the answers.' Consequently, the 71/100 accuracy in Fig. 6 should be interpreted as the accuracy of selecting the top-scoring answer according to the evaluators, not as evidence that the highest-quality answer is selected. The authors need to either state and validate an evaluator-reliability assumption or restrict the protocol's guarantee to robust score aggregation.
  2. [Section V-B-1, Fig. 6] The accuracy comparison against quorum-based voting is biased in favor of DecentLLMs by construction. The leader's answer in the baselines is assumed to be the ceil((Nw+1)/3)-th or ceil((Nw+1)/2)-th ranked answer, i.e., a deliberately mid- or low-ranked answer, rather than the answer an actual leader-election process would produce. Even if this is meant to model an underperforming leader, fixing the leader's rank at the lower end pre-selects the baselines to lose, and the reported gains (71 vs 64 vs 50 on 100 problems) are not accompanied by any confidence intervals or significance tests. A fair comparison should either implement a concrete leader-election mechanism, sweep the leader's rank over the full distribution, or report statistical uncertainty.
  3. [Section V-B-2, Fig. 7] The latency comparison is not end-to-end. The text states that DecentLLMs's consensus latency is measured from the point when evaluators all-to-all broadcast their evaluation results, which excludes worker answer generation and the evaluators' scoring phase, while the leader-driven baselines include multiple rounds of proposal and voting. Since LLM invocations dominate the runtime, the figure primarily shows that the measured portion of DecentLLMs is constant, not that the full protocol is faster. To support the 'faster consensus' claim, the authors should measure the total time from user prompt to final answer under the same conditions for all protocols.
  4. [Section III-A] The honest-majority condition is stated inconsistently. The formal assumption reads fw < floor((Nw-1)/2) and fe < floor((Ne-1)/2), but later the protocol relies on the GM tolerance bound f <= floor((n-1)/2) (Section III-C-3), and the experiment in Fig. 8 uses seven Byzantine evaluators among fifteen, which violates the stated condition (floor((15-1)/2)=7, so strict inequality allows at most six). The correct honest-majority condition is f < n/2; as written, the assumption is stronger than needed for odd n and inconsistent with the experimental evaluation. Please correct the definition and ensure the stated guarantee matches the threshold used in the experiments.
minor comments (4)
  1. [Section V-B-3] The reference 'Fig. V-B2' is not a valid figure reference; it should refer to Table I.
  2. [Section III-C-2 and Section V-A] The scoring scale is inconsistent: Section III-C-2 states each criterion is scored from 0 to 20, while Section V-A describes Byzantine evaluators assigning 'full scores (i.e., 100)' and Table I shows scores on a 0-100 scale. Please align the scoring-scale description with the experimental implementation.
  3. [Section V-B-3] There is a typo: 'the the next ranked answer' should read 'the next ranked answer'.
  4. [Section VII] In the conclusion, 'scores and ranking them' should be 'score and rank them'.

Circularity Check

1 steps flagged · score 3.0 of 10

Modest definitional circularity in the 'highest-quality answer' guarantee; the Byzantine-robust aggregation itself is self-contained.

  1. self definitional [Section III-C-3 (Eq. 1) and Section VII Conclusion]
    "After calculating the aggregated scores s_i for all answers, the evaluators identify the best answer a_best as the one with the highest robust score. ... This design enables faster consensus and ensures that the highest-quality answer is selected, even in the presence of Byzantine agents."

    Eq. (1) defines s_i = GM({A^k_i}), and the protocol defines a_best as argmax_i s_i. The conclusion then calls a_best 'the highest-quality answer.' If 'quality' means the evaluator-supplied scores, this is a tautological restatement of the selection rule, not a derived robustness property; if 'quality' means external correctness, no theorem connects GM aggregation to correctness. The paper's own Table I contradicts the external reading: incorrect w8 receives robust score 83.0 while correct w3 receives 55.6, and Section V-B-3 admits 'the evaluators failed to reliably score the answers.' Thus the headline guarantee reduces to the protocol's definition unless an unstated evaluator-reliability assumption is added.

full rationale

The core mechanism is self-contained: the geometric median (Eq. 1) is a fixed, standard robust estimator adopted from external literature [22], with no fitted parameter later renamed as a prediction. There is no load-bearing self-citation or imported uniqueness theorem; [19] and [20] are baselines, and [22] and [23] are external references. The single genuine circularity is in the quality claim: a_best is defined as the highest GM-aggregated evaluator score, and the conclusion re-labels that argmax as the 'highest-quality answer.' The paper's own Section V-B-3 and Table I show that this label can fail against external correctness, so the guarantee is conditional on an unstated evaluator-reliability assumption rather than a consequence of Byzantine-robust aggregation. Section III-B also asserts, without proof, that GM 'reliably supports the selection of the best available answer'; this is missing support rather than circular derivation. Separately, the quorum baselines are modeled with the leader at a fixed lower rank (ceil((Nw+1)/3) or ceil((Nw+1)/2)), so part of the reported accuracy margin is structural to the setup; this is a modeling/fairness concern, not a circular derivation. Overall, the protocol's real contribution is independent, but the headline guarantee is partly tautological, hence a moderate score of 3.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The protocol assumes synchronous communication, honest majorities in both groups, secure crypto, and reliable broadcast. The key unproven assumption is that LLM evaluator scores are meaningful quality indicators; the paper's own example contradicts this. No free parameters are fitted and no new entities are introduced.

assumptions (7)
  • domain assumption The network is synchronous with a bounded, predetermined maximum communication delay.
    Stated in Section III-A; all protocol timing and reliable broadcast arguments depend on this synchrony.
  • domain assumption Honest majority in both worker and evaluator groups: fw < floor((Nw-1)/2) and fe < floor((Ne-1)/2).
    Stated in Section III-A; the Byzantine resilience guarantee is conditional on these bounds.
  • domain assumption Standard cryptographic primitives (signatures, hashes) are secure and uncompromised.
    Stated in Section III-A; required for authentication and tie-breaking.
  • domain assumption A Byzantine reliable broadcast protocol exists and is used for all worker and evaluator messages.
    Invoked in Sections III-C-1 and III-C-3; the paper does not implement or evaluate this component.
  • standard math The geometric median tolerates up to f <= floor((n-1)/2) Byzantine vectors.
    Cited to [22] in Section III-C-3; used as the backbone of the score aggregation claim.
  • domain assumption LLM evaluator scores on the five criteria are meaningful indicators of answer quality.
    Implicit in Sections III-C-2 and V; the paper's own Table I shows scores can rank incorrect answers above correct ones.
  • domain assumption Honest workers do not generate answers that are overly anomalous beyond a known quality distribution.
    Stated in Section III-A; used to justify that honest worker answers are distinguishable from Byzantine sabotage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Byzantine-Robust Decentralized Coordination of LLM Agents." pith.science (2026). https://pith.science/paper/QWIVJKYO

@misc{pith2026250714928,
  author       = {Pith},
  title        = {Pith review of: Byzantine-Robust Decentralized Coordination of LLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QWIVJKYO}},
  note         = {Machine review of arXiv:2507.14928}
}
read the original abstract

Collaboration among multiple large language model (LLM) agents is a promising approach to overcome inherent limitations of single-agent systems, such as hallucinations and single points of failure. As LLM agents are increasingly deployed on open blockchain platforms, multi-agent systems capable of tolerating malicious (Byzantine) agents have become essential. Recent Byzantine-robust multi-agent systems typically rely on leader-driven coordination, which suffers from two major drawbacks. First, they are inherently vulnerable to targeted attacks against the leader. If consecutive leaders behave maliciously, the system repeatedly fails to achieve consensus, forcing new consensus rounds, which is particularly costly given the high latency of LLM invocations. Second, an underperforming proposal from the leader can be accepted as the final answer even when higher-quality alternatives are available, as existing methods finalize the leader's proposal once it receives a quorum of votes. To address these issues, we propose DecentLLMs, a novel decentralized consensus approach for multi-agent LLM systems, where worker agents generate answers concurrently and evaluator agents independently score and rank these answers to select the best available one. This decentralized architecture enables faster consensus despite the presence of Byzantine agents and consistently selects higher-quality answers through Byzantine-robust aggregation techniques. Experimental results demonstrate that DecentLLMs effectively tolerates Byzantine agents and significantly improves the quality of selected answers.

Figures

Figures reproduced from arXiv: 2507.14928 by the authors.

Figure 1
Figure 1. Motivating Examples: Leader-based consensus is vulnerable to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparisons with leader-based and leader-less approaches in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Operations of Existing Leader-Based Consensus in Byzantine-Robust [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: A protocol overview of DecentLLMs. who judge those answers and record the results on-chain. When workers generate and submit their responses, evaluators attempt to build a block that contains their assessments and the corresponding rewards. The evaluator who first crea…
Figure 5
Figure 5. Figure 5: Prompts used in DecentLLMs. Role-playing prompting is applied throughout, with additional chain-of-thought prompting for workers and few-shot [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Accuracy comparisons of DecentLLMs and other quorum-based voting [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Number of Byzantine Evaluators. With eight honest evaluators, [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RED-SEGA:Resilient Decentralized Stochastic Proximal Optimization with Gradient Sketching over Time-Varying Networks

    math.OC 2026-07 conditional novelty 6.0 of 10

    RED-SEGA achieves Byzantine-resilient linear convergence for non-decomposable SRM via gradient sketching and norm-penalized aggregation over time-varying networks.

  2. Engineering Trustworthy Agentic AI for Critical Systems

    cs.AI 2026-07 conditional novelty 5.0 of 10

    A survey claiming that agentic AI trustworthiness is a single cross-domain problem and outlining a framework for graded, certifiable assurance.

  3. Blockchain Empowered Trustworthy Agent Networks: Foundations, Taxonomy, and Future Directions

    cs.CR 2026-08 conditional novelty 4.0 of 10

    A survey proposing a five-dimensional taxonomy of trust crises in open AI agent networks and analyzing blockchain's role as a shared trust infrastructure.

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages · cited by 3 Pith papers

  1. [1]

    (2024, May) Gpt-4o

    OpenAI. (2024, May) Gpt-4o. [Online]. Available: https://openai.com/index/hello-gpt-4o//

    work page 2024

  2. [2]

    (2025, May) Qwen3

    Qwen3. (2025, May) Qwen3. [Online]. Available: https://ollama.com/library/qwen3/

    work page 2025

  3. [3]

    (2025, May) Claude ai

    Anthropic. (2025, May) Claude ai. [Online]. Available: https://claude.ai/

    work page 2025

  4. [4]

    (2025, May) Llama

    Meta. (2025, May) Llama. [Online]. Available: https://www.llama.com/

    work page 2025

  5. [5]

    (2025, May) Google gemini

    Google. (2025, May) Google gemini. [Online]. Available: https://gemini.google.com/

    work page 2025

  6. [6]

    (2025, May) Deepseek

    DeepSeek. (2025, May) Deepseek. [Online]. Available: https://www.deepseek.com/

    work page 2025

  7. [7]

    Improving factuality and reasoning in language models through multiagent debate,

    Y . Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch, “Improving factuality and reasoning in language models through multiagent debate,” in Proceedings of the 41st International Conference on Machine Learn- ing, ser. ICML’24. JMLR.org, 2024

    work page 2024

  8. [8]

    Human-like summarization evaluation with chatgpt,

    M. Gao, J. Ruan, R. Sun, X. Yin, S. Yang, and X. Wan, “Human-like summarization evaluation with chatgpt,” 2023. [Online]. Available: https://arxiv.org/abs/2304.02554

    arXiv 2023

Show all 41 references
  1. [9]

    Is multi-agent debate (mad) the silver bullet? an empirical analysis of mad in code summarization and translation,

    J. Chun, Q. Chen, J. Li, and I. Ahmed, “Is multi-agent debate (mad) the silver bullet? an empirical analysis of mad in code summarization and translation,” 2025. [Online]. Available: https://arxiv.org/abs/2503.12029

    2025 arXiv

  2. [10]

    DA-code: Agent data science code generation benchmark for large language models,

    Y . Huang, J. Luo, Y . Yu, Y . Zhang, F. Lei, Y . Wei, S. He, L. Huang, X. Liu, J. Zhao, and K. Liu, “DA-code: Agent data science code generation benchmark for large language models,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , Y...

    2024

  3. [11]

    Llmsmartsec: Smart contract security auditing with llm and annotated control flow graph,

    V . Mothukuri, R. M. Parizi, and J. L. Massa, “Llmsmartsec: Smart contract security auditing with llm and annotated control flow graph,” in 2024 IEEE International Conference on Blockchain (Blockchain) , 2024, pp. 434–441

    2024

  4. [12]

    Blockchain large language models,

    Y . Gai, L. Zhou, K. Qin, D. Song, and A. Gervais, “Blockchain large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2304.12749

    2023 arXiv

  5. [13]

    Sequential diagnosis with language models,

    H. Nori, M. Daswani, C. Kelly, S. Lundberg, M. T. Ribeiro, M. Wilson, X. Liu, V . Sounderajah, J. Carlson, M. P. Lungren, B. Gross, P. Hames, M. Suleyman, D. King, and E. Horvitz, “Sequential diagnosis with language models,” 2025. [Online]. Available: https://arxiv.org/abs/2506.22405

    2025 arXiv

  6. [14]

    Should we be going mad? a look at multi-agent debate strategies for llms,

    A. Smit, N. Grinsztajn, P. Duckworth, T. D. Barrett, and A. Pretorius, “Should we be going mad? a look at multi-agent debate strategies for llms,” in Proceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

    2024

  7. [15]

    More agents is all you need,

    junyou li, Q. Zhang, Y . Yu, Q. FU, and D. Ye, “More agents is all you need,” Transactions on Machine Learning Research , 2024. [Online]. Available: https://openreview.net/forum?id=bgzUSZ8aeg

    2024

  8. [16]

    Beyond self-talk: A communication-centric survey of llm-based multi-agent systems,

    B. Yan, X. Zhang, L. Zhang, L. Zhang, Z. Zhou, D. Miao, and C. Li, “Beyond self-talk: A communication-centric survey of llm-based multi-agent systems,” 2025. [Online]. Available: https://arxiv.org/abs/2502.14321

    2025 arXiv

  9. [17]

    When one llm drools, multi-llm collaboration rules,

    S. Feng, W. Ding, A. Liu, Z. Wang, W. Shi, Y . Wang, Z. Shen, X. Han, H. Lang, C.-Y . Lee, T. Pfister, Y . Choi, and Y . Tsvetkov, “When one llm drools, multi-llm collaboration rules,” 2025. [Online]. Available: https://arxiv.org/abs/2502.04506

    2025 arXiv

  10. [18]

    Large language model based multi-agents: A survey of progress and challenges,

    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,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson...

    2024 doi

  11. [19]

    Blockagents: Towards byzantine-robust llm-based multi-agent coordination via blockchain,

    B. Chen, G. Li, X. Lin, Z. Wang, and J. Li, “Blockagents: Towards byzantine-robust llm-based multi-agent coordination via blockchain,” in Proceedings of the ACM Turing Award Celebration Conference - China 2024, ser. ACM-TURC ’24. New York, NY , USA: Association for Computing M...

    2024

  12. [20]

    A weighted byzantine fault tolerance consensus driven trusted multiple large language models network,

    H. Luo, G. Sun, Y . Liu, D. Zhao, D. Niyato, H. Yu, and S. Dustdar, “A weighted byzantine fault tolerance consensus driven trusted multiple large language models network,” 2025. [Online]. Available: https://arxiv.org/abs/2505.05103

    2025 arXiv

  13. [21]

    Hotstuff: Bft consensus with linearity and responsiveness,

    M. Yin, D. Malkhi, M. K. Reiter, G. G. Gueta, and I. Abraham, “Hotstuff: Bft consensus with linearity and responsiveness,” in Proceedings of the 2019 ACM Symposium on Principles of Distributed Computing, ser. PODC ’19. New York, NY , USA: Association for Computing Machinery, 2...

    2019

  14. [22]

    Byzantine machine learning: A primer,

    R. Guerraoui, N. Gupta, and R. Pinot, “Byzantine machine learning: A primer,” ACM Comput. Surv. , vol. 56, no. 7, Apr. 2024. [Online]. Available: https://doi.org/10.1145/3616537

    2024 doi

  15. [23]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu, “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” ACM Trans. Inf. Syst. , vol. 43, no. 2, Jan. 2025. [Online]. Available...

    2025 doi

  16. [24]

    (2025, May) io.net

    IO.Net. (2025, May) io.net. [Online]. Available: https://io.net/

    2025

  17. [25]

    T. A. N. Authors. (2025, May) Akash network. [Online]. Available: https://akash.network//

    2025

  18. [26]

    T. B. Authors. (2025, May) Bittensor. [Online]. Available: https://bittensor.com/

    2025

  19. [27]

    A survey of recent backdoor attacks and defenses in large language models,

    S. Zhao, M. Jia, Z. Guo, L. Gan, X. XU, X. Wu, J. Fu, F. Yichao, F. Pan, and A. T. Luu, “A survey of recent backdoor attacks and defenses in large language models,” Transactions on Machine Learning Research , 2025, survey Certification. [Online]. Available: https://openreview....

    2025

  20. [28]

    A survey on trustworthy llm agents: Threats and countermeasures,

    M. Yu, F. Meng, X. Zhou, S. Wang, J. Mao, L. Pang, T. Chen, K. Wang, X. Li, Y . Zhang, B. An, and Q. Wen, “A survey on trustworthy llm agents: Threats and countermeasures,” 2025. [Online]. Available: https://arxiv.org/abs/2503.09648

    2025 arXiv

  21. [29]

    Randomized vs. deterministic? practical randomized synchronous BFT in expected constant time,

    X. Zhang, B. Huang, S. Duan, and H. Zhang, “Randomized vs. deterministic? practical randomized synchronous BFT in expected constant time,” Cryptology ePrint Archive, Paper 2025/816, 2025. [Online]. Available: https://eprint.iacr.org/2025/816

    2025

  22. [30]

    Ma- chine learning with adversaries: Byzantine tolerant gradient descent,

    P. Blanchard, E. M. El Mhamdi, R. Guerraoui, and J. Stainer, “Ma- chine learning with adversaries: Byzantine tolerant gradient descent,” in Advances in Neural Information Processing Systems , I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. G...

    2017

  23. [31]

    The hidden vul- nerability of distributed learning in byzantium,

    E. M. El Mhamdi, R. Guerraoui, and S. Rouault, “The hidden vul- nerability of distributed learning in byzantium,” in Proceedings of the 35th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. PMLR, ...

    2018

  24. [32]

    (2025, May) Geometric median

    wikipedia. (2025, May) Geometric median. [Online]. Available: https://en.wikipedia.org/wiki/Geometric median/

    2025

  25. [33]

    P. S. Foundation. (2024, May) Python 3.11.11. [Online]. Available: https://www.python.org/downloads/release/python-31111/

    2024

  26. [34]

    Mmlu-pro: A more robust and chal- lenging multi-task language understanding benchmark,

    Y . Wang, X. Ma, G. Zhang, Y . Ni, A. Chandra, S. Guo, W. Ren, A. Arulraj, X. He, Z. Jiang et al., “Mmlu-pro: A more robust and chal- lenging multi-task language understanding benchmark,” arXiv preprint arXiv:2406.01574, 2024

    2024 arXiv

  27. [35]

    Practical byzantine fault tolerance,

    M. Castro, B. Liskov et al. , “Practical byzantine fault tolerance,” in OSDI, vol. 99, 1999, pp. 173–186

    1999

  28. [36]

    Narwhal and tusk: A dag-based mempool and efficient bft consensus,

    G. Danezis, L. Kokoris-Kogias, A. Sonnino, and A. Spiegelman, “Narwhal and tusk: A dag-based mempool and efficient bft consensus,” in Proceedings of the Seventeenth European Conference on Computer Systems, ser. EuroSys ’22. New York, NY , USA: Association for Computing Machine...

    2022

  29. [37]

    Examining inter-consistency of large language models collaboration: An in- depth analysis via debate,

    K. Xiong, X. Ding, Y . Cao, T. Liu, and B. Qin, “Examining inter-consistency of large language models collaboration: An in- depth analysis via debate,” in Findings of the Association for Computational Linguistics: EMNLP 2023 . Association for Computational Linguistics, 2023, p...

    2023 doi

  30. [38]

    Chainlink 2.0: Next steps in the evolution of decentralized oracle networks,

    L. Breidenbach, C. Cachin, B. Chan, A. Coventry, S. Ellis, A. Juels, F. Koushanfar, A. Miller, B. Magauran, D. Moroz et al., “Chainlink 2.0: Next steps in the evolution of decentralized oracle networks,” Chainlink Labs, vol. 1, pp. 1–136, 2021

    2021

  31. [39]

    (2025, May) Supra: A faster, better web3 experience for everyone

    Supra. (2025, May) Supra: A faster, better web3 experience for everyone. [Online]. Available: https://supra.com/

    2025

  32. [40]

    (2025, May) Ai-powered oracles: Bridging blockchains with smarter data

    kava. (2025, May) Ai-powered oracles: Bridging blockchains with smarter data. [Online]. Available: https://www.kava.io/news/ai-powered- oracles/

    2025

  33. [41]

    (2025, May) Redstone: The modular blockchain oracle

    RedStone. (2025, May) Redstone: The modular blockchain oracle. [Online]. Available: https://www.redstone.finance/

    2025

Pith tools

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