The specifics:
The method employs a two-phase process: AI "workers" solve problems and evaluate their own answers, and the best contribution is then chosen in a tournament-style bracket.
Nomos received eight flawless problem scores last year, placing it second out of almost 4,000 human participants.
Additionally, Nous published and made available as open-source a reasoning harness, which is orchestration code that controls the model's problem-solving process.
When Qwen3 was run through the same harness and setup, the score was only 24/120, indicating that model training rather than the harness was responsible for the increases.
Nomos received eight flawless problem scores last year, placing it second out of almost 4,000 human participants.
Additionally, Nous published and made available as open-source a reasoning harness, which is orchestration code that controls the model's problem-solving process.
When Qwen3 was run through the same harness and setup, the score was only 24/120, indicating that model training rather than the harness was responsible for the increases.
Even basic math issues were a barrier for the best AI systems not too long ago, but today a tiny, open model is passing a famously challenging test. The entire industry is poised for an AI-driven boom, with Nomos, AI helping solve intractable problems, and labs bringing in gold medal-winning math models.