DiPO disentangles samples via perplexity to enable fine-grained exploration-exploitation in RLVR, using bidirectional reward allocation to improve LLM performance on mathematical reasoning and function calling.
Instead, we utilize max-PPL reward and max-PPL penalty as training reward, respectively, and recorded the changes in model entropy
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
DiPO: Disentangled Perplexity Policy Optimization for Fine-grained Exploration-Exploitation Trade-Off
DiPO disentangles samples via perplexity to enable fine-grained exploration-exploitation in RLVR, using bidirectional reward allocation to improve LLM performance on mathematical reasoning and function calling.