Mechanism

  • Absolute Zero Reasoner (AZR): Eliminates human-curated prompt datasets entirely by training a single model in two roles coupled via a Python code executor: a Proposer (rewarded via RL for generating verifiable deduction/abduction/induction tasks of intermediate learnability 0<rˉsolver<10 < \bar{r}_{\text{solver}} < 1) and a Solver (rewarded for solving the proposed tasks).
    • Test-Time Reinforcement Learning (TTRL): Performs online GRPO on unlabelled test prompts by using majority-voted consensus answers across GG rollouts as pseudo-verifiable rewards.

References

  1. Absolute Zero: Reinforced Self-play Reasoning with Zero Data ↗

    Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue et al. · 2025

    All 11 authors

    Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue, Matthieu Lin, Shenzhi Wang, Qingyun Wu, Zilong Zheng, Gao Huang

  2. TTRL: Test-Time Reinforcement Learning ↗

    Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu et al. · 2025

    All 16 authors

    Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, Biqing Qi, Youbang Sun, Zhiyuan Ma, Lifan Yuan, Ning Ding, Bowen Zhou