Collaborative RLVR for Robust Reasoning

Reinforcement learning with verifiable rewards is a powerful way to train LLMs for mathematical reasoning, but it often produces brittle behavior: models exploit dataset patterns, generate unfaithful reasoning, and fail under simple rephrasings or distribution shifts. The idea is that by making RLVR collaborative - i.e. having two models solve independently and then discuss before answering, we force reasoning to be externalized/challenged/defended, while keeping rewards strictly correctness-based. If a solution has to convince another agent with a different inductive bias, shallow heuristics should break, while genuinely robust reasoning survives.