G-Zero

G-Zero: Self-Play for Open-Ended Generation from Zero Data

Chengsong Huang, Haolin Liu et al. · arXiv 2026

Read paper · arXiv:2605.09959

Drive open-ended self-play with hint-induced probability shifts.

What you will implement

G-Zero Hint-Delta; G-Zero Preference Curation

• 109. G-Zero Hint-Delta

  G-Zero Hint-Delta

  Align same-answer token probabilities to compute Hint-δ and proposer rewards.

• 110. G-Zero Preference Curation

  G-Zero Preference Curation

  Filter by the full-pool median and build preference pairs with question-only prompts.