Off-policy GRPO, which estimates advantages from a slightly stale policy, matches or improves on-policy GRPO on math reasoning tasks and is supported by a new policy-improvement lower bound.
The models that we used wereQwen/Qwen2.5-0.5B-Instruct (Apache REVISITING GROUP RELATIVE POLICY OPTIMIZATION 15 2.0 license) by Yang et al
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
method 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
method 1polarities
use method 1representative citing papers
citing papers explorer
-
Revisiting Group Relative Policy Optimization: Insights into On-Policy and Off-Policy Training
Off-policy GRPO, which estimates advantages from a slightly stale policy, matches or improves on-policy GRPO on math reasoning tasks and is supported by a new policy-improvement lower bound.