Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fortytwo: Swarm Inference with Peer-Ranked Consensus

Vladyslav Larin, Ihor Naumenko, Aleksei Ivashov, Ivan Nikitin, Alexander Firsov

TL;DR

Fortytwo addresses the scalability bottleneck of centralized AI by distributing inference across a swarm of heterogeneous nodes that jointly generate and rank outputs. The core method combines distributed pairwise ranking via a Bradley-Terry–style aggregation with multi-token reasoning, reputation-weighted consensus, compute-stake Sybil defenses, and on-chain coordination. Key results show substantial gains over majority voting (e.g., GPQA Diamond from 68.69% to 85.90%) and exceptional robustness to adversarial and noisy prompts (0.12% degradation vs 6.20% for monolithic models) across six benchmarks. The work demonstrates that decentralized, meritocratic swarms can deliver high-quality, secure AI inference while democratizing access and maintaining practical deployability, offering a blueprint for collective intelligence in AI systems.

Abstract

As centralized AI hits compute ceilings and diminishing returns from ever-larger training runs, meeting demand requires an inference layer that scales horizontally in both capacity and capability. We present Fortytwo, a novel protocol that leverages swarm intelligence principles and distributed pairwise ranking consensus to achieve superior performance in AI inference. Our approach reimagines collaboration among AI nodes using swarm inference: a peer-ranked, reputation-weighted consensus across heterogeneous models that surfaces the highest-quality responses. Using pairwise ranking with a custom Bradley-Terry-style aggregation model, we demonstrate that swarm inference substantially outperforms majority voting, achieving 85.90% on GPQA Diamond versus 68.69% for majority voting with the same model set - an improvement of +17.21 percentage points (approximately +25.1% relative). The protocol incorporates on-chain reputation so node influence adapts to demonstrated accuracy over time, yielding a meritocratic consensus that filters low-quality or malicious participants. To resist Sybil attacks, Fortytwo employs proof-of-capability in its consensus: nodes must successfully complete calibration/test requests and stake reputation to enter ranking rounds, making multi-identity attacks economically unattractive while preserving openness. Across six challenging benchmarks, including GPQA Diamond, LiveCodeBench, and AIME, our evaluation indicates higher accuracy and strong resilience to adversarial and noisy free-form prompting (e.g., prompt-injection degradation of only 0.12% versus 6.20% for a monolithic single-model baseline), while retaining practical deployability. Together, these results establish a foundation for decentralized AI systems - democratizing access to high-quality inference through collective intelligence without sacrificing reliability or security.

Fortytwo: Swarm Inference with Peer-Ranked Consensus

TL;DR

Fortytwo addresses the scalability bottleneck of centralized AI by distributing inference across a swarm of heterogeneous nodes that jointly generate and rank outputs. The core method combines distributed pairwise ranking via a Bradley-Terry–style aggregation with multi-token reasoning, reputation-weighted consensus, compute-stake Sybil defenses, and on-chain coordination. Key results show substantial gains over majority voting (e.g., GPQA Diamond from 68.69% to 85.90%) and exceptional robustness to adversarial and noisy prompts (0.12% degradation vs 6.20% for monolithic models) across six benchmarks. The work demonstrates that decentralized, meritocratic swarms can deliver high-quality, secure AI inference while democratizing access and maintaining practical deployability, offering a blueprint for collective intelligence in AI systems.

Abstract

As centralized AI hits compute ceilings and diminishing returns from ever-larger training runs, meeting demand requires an inference layer that scales horizontally in both capacity and capability. We present Fortytwo, a novel protocol that leverages swarm intelligence principles and distributed pairwise ranking consensus to achieve superior performance in AI inference. Our approach reimagines collaboration among AI nodes using swarm inference: a peer-ranked, reputation-weighted consensus across heterogeneous models that surfaces the highest-quality responses. Using pairwise ranking with a custom Bradley-Terry-style aggregation model, we demonstrate that swarm inference substantially outperforms majority voting, achieving 85.90% on GPQA Diamond versus 68.69% for majority voting with the same model set - an improvement of +17.21 percentage points (approximately +25.1% relative). The protocol incorporates on-chain reputation so node influence adapts to demonstrated accuracy over time, yielding a meritocratic consensus that filters low-quality or malicious participants. To resist Sybil attacks, Fortytwo employs proof-of-capability in its consensus: nodes must successfully complete calibration/test requests and stake reputation to enter ranking rounds, making multi-identity attacks economically unattractive while preserving openness. Across six challenging benchmarks, including GPQA Diamond, LiveCodeBench, and AIME, our evaluation indicates higher accuracy and strong resilience to adversarial and noisy free-form prompting (e.g., prompt-injection degradation of only 0.12% versus 6.20% for a monolithic single-model baseline), while retaining practical deployability. Together, these results establish a foundation for decentralized AI systems - democratizing access to high-quality inference through collective intelligence without sacrificing reliability or security.

Paper Structure

This paper contains 138 sections, 23 equations, 7 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Multidisciplinary Foundations
  3. Key Innovations and Contributions
  4. Paper Organization
  5. Related Work
  6. Cryptographic Approaches to Verifiable AI
  7. Cryptographic Approaches to Verifiable AI
  8. Zero-Knowledge Machine Learning
  9. Homomorphic Encryption
  10. Optimistic and Economic Security Models
  11. Optimistic Machine Learning
  12. Proof of Quality Mechanisms
  13. Distributed Consensus and Byzantine Fault Tolerance
  14. Classical Byzantine Fault Tolerance
  15. Modern BFT Innovations
  16. ...and 123 more sections

Figures (7)

  • Figure 1: Benchmark performance comparison across models.
  • Figure 2: Modular architecture of a self-supervised inference showing the four key components and their interactions
  • Figure 3: Evolution of node reputation scores over time showing different performance patterns
  • Figure 4: Benchmark performance comparison across models. Fortytwo achieves state-of-the-art results on LiveCode (84.4%), MATH-500 (99.6%), AIME 2024 (100%), and AIME 2025 (96.66%), demonstrating superior performance on challenging reasoning and coding benchmarks.
  • Figure 5: Performance scaling with swarm size showing rapid improvement and convergence
  • ...and 2 more figures