d-TreeRPO uses tree rollouts for fine-grained verifiable rewards and time-scheduled self-distillation to reduce probability estimation gaps in diffusion LLMs, delivering substantial gains on Sudoku, Countdown, GSM8K, and Math500 benchmarks.
Since ptrue d is the expectation of qτ(d,σ) (σ) over σ∼ Q , reducing the typical-path devi- ation (smaller ϵd,δ) makes ˆpd a more reliable proxy when approximatingp true d
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models
d-TreeRPO uses tree rollouts for fine-grained verifiable rewards and time-scheduled self-distillation to reduce probability estimation gaps in diffusion LLMs, delivering substantial gains on Sudoku, Countdown, GSM8K, and Math500 benchmarks.