Standard diffusion language models suffer from globally unconstrained decoding that accumulates early errors, while block diffusion sacrifices semantic coherence at hard block boundaries. We propose a dynamic wavefront mechanism inspired by physical wave propagation: an active token frontier expands outward from already-decoded positions, with adaptive Expand and confidence-based Prune steps. On GSM8K, HumanEval, and three additional math/code reasoning benchmarks, WavefrontDiffusion surpasses BlockDiffusion across the board while keeping compute on par. We introduce the MHCO metric to quantify boundary-induced reasoning violations and show consistent gains in BERTScore semantic consistency.

Multi-turn industrial sales dialogues mix sparse long-horizon business signals (conversion) with dense short-horizon language constraints (fluency, compliance), making joint optimization unstable and prone to reward hacking. We propose Dual-Horizon Credit Assignment (DuCA), whose HIAN mechanism applies decoupled normalization to turn-level and session-level advantages so that high-variance commercial rewards no longer suppress gradients on fine-grained linguistic signals. On a high-fidelity simulated commercial environment, DuCA improves conversion by 6.82% over GRPO while cutting multi-turn repetition by 82.28%.

We bridge on-policy RL exploitation and off-policy imitation at the token level via an asymmetric distillation scheme.