Continuous diffusion language models struggle with structured generation under hard, variable-length templates. We identify two failure modes: Boundary Attention Drift, in which head-region attention bleeds into padded tails under block-mode training, and Vocab Projection Drift, in which latent-to-token rounding over-projects to low-frequency tokens. The two modes dissociate cleanly across Chinese ci, English sonnets, and E2E NLG: the former is specific to block-mode training, whereas the latter appears in all three datasets regardless of padding regime, isolating padding regime and rounding rule as orthogonal sources of error. To address structural correctness, we adapt constrained decoding to diffusion rounding through a position-template primitive, the V_i rule, instantiated on continuous-diffusion logits as Length-Conditional Voronoi Rounding (LCVR). The same primitive applied to a small autoregressive baseline reaches identical zero-violation rates, showing that structural correctness is delivered by the template, not by the underlying model family. A contribution more readily realised with diffusion backbones is instead cross-domain backbone transfer: initialising a target-domain model from a structurally compatible source-domain backbone yields order-of-magnitude reductions in training steps, while a controlled negative-transfer arm fails to reproduce the speed-up on structural metrics; an attention-span probe localises the transferable component to the source backbone’s long-range attention span rather than to its padding regime or corpus size. We further report a calibrated wallclock analysis showing that the dominant cost of diffusion inference is recoverable with standard step-reduction techniques. Together, these results recast several previously scattered structured-generation failures of continuous diffusion as one consistent, transferable, and fixable phenomenon.

Group-relative policy optimization (GRPO, DAPO, ARPO) has become the dominant recipe for RL fine-tuning of reasoning LLMs, yet in the resource-constrained regime most practitioners train in (a single GPU, small batches) it consistently suffers from rapid entropy collapse that kills exploration, together with sparse credit assignment in long rollouts, which is most extreme under multi-turn tool use but already present in long single-turn chain-of-thought. We address both failure modes with JANUSPO, a drop-in GRPO modification that introduces asymmetric PPO clipping 𝜖₊<𝜖_- to tolerate larger downward than upward updates, pairs it with an asymmetric KL schedule 𝛽_pos<𝛽_neg that doubles as a free entropy regulator without any explicit upper bound, and redistributes the terminal reward through a turn-weighted decomposition w_t(n) that generalises naturally from single-turn CoT to multi-turn rollouts. On Qwen2.5-3B-Instruct / GSM8K, JANUSPO improves from 74.8% to 84.3±0.2% (mean±std over 2 seeds) across 14 controlled ablations, and the same recipe lifts 7B from 85.3% to 93.6% with no hyper-parameter retuning. On multi-turn DeepSearch (103 GAIA + 100 HLE), SFT-warmup-then-JANUSPO is the only RL recipe whose paired comparison against SFT survives best@8 outlier verification, achieving strict dominance on all four reward and F1 columns while a matched-init GRPO baseline’s apparent F1 wins collapse to 0%. Beyond this main result, we document a systematic negative finding: SFT warm-up is a necessary precondition for GRPO-style multi-turn RL on format-unaware base models.