AGPO applies asymmetric negative-dominant reinforcement and variance-scaled group advantages in RLVR to preserve base model exploration while boosting accuracy and pass@k on math benchmarks and industrial ad relevance data.
Clearly, this is possible since we can always find two distinct elements ofA 1 whose sum isnforn≥15
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
AGPO: Asymmetric Group Policy Optimization for Verifiable Reasoning and Search Ads Relevance at JD
AGPO applies asymmetric negative-dominant reinforcement and variance-scaled group advantages in RLVR to preserve base model exploration while boosting accuracy and pass@k on math benchmarks and industrial ad relevance data.