DEPO combines diversity/influence/difficulty-aware offline data selection with an entropy-based online rollout filter and replay, matching full-data GRPO performance with 20% data and 40% rollouts.
This objective aims to select a subset Y that is both diverse (as captured by det(SY)) and influential (as promoted by the product of weights ∏i∈Y wi)
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
-
Towards High Data Efficiency in Reinforcement Learning with Verifiable Reward
DEPO combines diversity/influence/difficulty-aware offline data selection with an entropy-based online rollout filter and replay, matching full-data GRPO performance with 20% data and 40% rollouts.