R-Zero

R-Zero: Self-Evolving Reasoning LLM from Zero Data

Chengsong Huang, Wenhao Yu et al. · ICLR 2026

Read paper · arXiv:2508.05004

Co-evolve a text challenger and solver.

What you will implement

R-Zero Challenger Rewards; R-Zero Pseudo-Labels and Curriculum

• 107. R-Zero Challenger Rewards

  R-Zero Challenger Rewards

  Implement triangular uncertainty, batch repetition penalty and format gating.

• 108. R-Zero Pseudo-Labels and Curriculum

  R-Zero Pseudo-Labels and Curriculum

  Count majority answers, filter intermediate consistency and score against frozen pseudo-labels.