GHPO adaptively adds partial ground-truth solution hints to prompts the model cannot currently solve, and claims average gains of about 4-5% on math benchmarks.
OpenAI o3 and o4-mini system card
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
REJECT 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
GHPO: Adaptive Guidance for Stable and Efficient LLM Reinforcement Learning
GHPO adaptively adds partial ground-truth solution hints to prompts the model cannot currently solve, and claims average gains of about 4-5% on math benchmarks.