Fortytwo scores 85.9% on GPQA Diamond by letting models judge each other

Large language models have now reached the level of an advanced expert on a wide range of tasks. Raising their quality and reliability further in decentralized use leaves only one option — making them substantially larger — and that is expensive and economically inefficient.

There is another way out: AI swarms — systems built from models that differ in architecture, knowledge and capability, each of which answers every query. Collective decisions, after all, often beat individual ones. The catch is that the result of a multi-agent majority vote stagnates quickly, giving up its robustness and performance advantage over a single model. Today's multi-agent systems also lean on large LLMs and usually work only in narrow scenarios.

What's the problem?

Distributed, decentralized swarm AI networks offer privacy, transparency, lower latency and a diversity of models. But they are hard to build, because of network latency and the absence of trust between participants. On top of that, like public blockchains, such networks suffer from Sybil attacks and from economic attacks, which makes them less dependable in practice — especially in open AI networks that anyone is free to join.

Cryptographic protocols such as zero-knowledge proofs or proof of generation let a user obtain an answer without trusting any single node, but they are expensive. What is needed is a cheaper protocol that works across all tasks while delivering the quality of leading models, the transparency, privacy and fault tolerance the setting demands, and resistance to Byzantine attacks.

Long live the blockchain

The researchers introduce Fortytwo, a fault-tolerant system for running an AI swarm built on pairwise comparisons, a tournament ranking and compute economics. The authors give every participant in the network a dual role: on one side it acts as an AI that supplies answers, on the other as a judge that measures the quality of its peers' solutions on each task through pairwise comparisons.

Aggregating those judgements rests on the classic probabilistic Bradley-Terry model, widely used to analyse sports competitions, which reconstructs a global ranking of the proposed solutions from pairwise comparisons. The model naturally accommodates factors that capture the relative competence of different agents' opinions, and constructive or destructive behaviour by users is reflected accordingly in each participant's individual reputation.

The reputation updates themselves are driven by how consistent and reliable a given user's votes are on each task. Every prospective judge writes a short chain of thought (50–100 tokens) on a "cryptographically" selected pair of answers produced by other users, filtered so that no conflict of interest is possible.

The global ranking is updated with a reputation-weighted Bradley-Terry estimate, which makes the voting more consistent. The reasoning chains guarantee that a judge's assessment is transparent and can be examined for the causes of isolated or systematic errors and abuse. The metadata can live on the blockchain, keeping a list of bad-faith participants and making reputation both immutable and easy to work with.

Large AI models today generally need a great deal of compute. That makes it possible to borrow the compute-reputation idea from decentralized and blockchain networks — an end-to-end "compute-stake" — to account for what a participant has put in and what they have earned. To win a vote in how solutions are ranked, participants have to demonstrate their compute capability through "proof-of-capability", tested on a range of benchmark tasks in the domains they care about, such as mathematics or coding.

Weighting a user by competence is consistent both with social choice theory and with analysis of real user-rating data. It makes Sybil attacks economically pointless, since every Sybil has to prove its ability on world-class tasks — which is beyond the overwhelming majority.

Experiments

The experiments measured Fortytwo's performance across a range of tasks. On GPQA Diamond, Fortytwo showed a substantial gain over plain majority voting: 85.90% against 68.69% (a relative gain of about 25.1%). On LiveCodeBench it reached 84.4%, on MATH-500 99.6%, and on the American Invitational Mathematics Examination (AIME) problem sets for 2024 and 2025, 100% and 96.66% respectively.

Takeaways

Broadly, the quality gain from growing the swarm keeps paying off up to roughly 30 participants. The important part: Fortytwo beats plain majority voting consistently, at every swarm size tested.

Fortytwo therefore shows that collective reasoning across many independent models can outperform even the largest monolithic AI models. A swarm of neural networks becomes not merely an alternative to centralized systems but a new way of organizing computational intelligence — open, resilient and economically balanced.

The future of artificial intelligence may well lie in architectures of this kind: not in a single superintelligent giant, but in a network of interacting minds, where truth emerges from many judgements.