Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning

Agon proposes competitive cross-model reinforcement learning in which two models solve the same problem while alternately drafting, reading, and competing against each other. Each model is rewarded for outperforming a rival that has seen its reasoning, creating implicit pressure to improve the trace without process labels or a learned reward model. The abstract reports that, on the hard split of DeepMath with Qwen3, Agon doubled GRPO pass@1, delivering roughly eight times the gain of an untrained Mixture-of-Agents baseline. The ordering reportedly replicated on competitive-programming code and across Qwen3.5 and Gemma 4.