Co-RL unlocks reasoning through diverse peers

Co-RL’s strongest idea is that diversity can serve as a scalable substitute for increasingly expensive reward supervision.

• –Heterogeneous model families, sizes, and prompt phrasing create less-correlated errors
• –The framework avoids shared parameters and external judges, relying on cross-agent majority-vote rewards
• –Reported gains reach 3.0–8.6% across seven text benchmarks and 2.3–7.2% across four multimodal benchmarks
• –It addresses a key weakness of self-rewarding RL: feedback loops that homogenize behavior and trigger training collapse
• –The open-source implementation makes the approach testable, though multi-model RL remains compute-intensive