A length-weighted reward baseline plus exact on-policy updates improves stability and diversity in LLM reinforcement learning for math, with small performance gains over GRPO.
American invitational mathematics examination - aime
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
On-Policy RL with Optimal Reward Baseline
A length-weighted reward baseline plus exact on-policy updates improves stability and diversity in LLM reinforcement learning for math, with small performance gains over GRPO.