LAVE introduces a lookahead‑then‑verify constrained decoding technique that dramatically raises syntax correctness for diffusion language models across code, JSON, and SMILES generation, improves functional correctness, and adds only minimal inference overhead.
The paper presents d‑OPSD, the first on‑policy self‑distillation framework for diffusion large language models that eliminates reference solutions and extra teacher models, using only one‑tenth of RL steps while achieving equal or superior reasoning performance and markedly higher training efficiency, as demonstrated on multiple math‑reasoning benchmarks.
A comprehensive evaluation shows that while diffusion language models achieve higher generation speed through parallel decoding, they suffer from severe causal reasoning and formatting deficiencies, lagging far behind autoregressive models on embodied and tool‑calling agent tasks.
Recent work from West Lake University's MAPLE Lab introduces a diffusion‑based “Divergent Thought Chain” that treats each intermediate denoising step of a diffusion language model as a reasoning step, using result‑based reinforcement learning to optimize non‑linear token generation and achieving state‑of‑the‑art performance on math and code tasks.
