Eight DLMs benchmarked on reasoning, coding, and translation tasks reveal inference-step trade-offs

The study trains and tests modern DLMs under matched conditions on tasks including GSM8K, HumanEval, and WMT translation. Controlled ablations isolate generation hyperparameters, showing that 50-step schedules close 70 percent of the gap to autoregressive baselines on knowledge recall yet require 3.2 times more FLOPs at equivalent context length. Block-wise unmasking improves coherence on structured outputs while increasing variance on open-ended reasoning.

Results demonstrate DLMs achieve lower perplexity than AR models at fixed inference budgets above 128 steps but degrade sharply below 16 steps on coding accuracy. Parallel refinement enables full-sequence updates unavailable to next-token predictors, yet exposes sensitivity to context truncation beyond 2k tokens. Smaller matched models confirm the pattern holds independent of scale.

The analysis positions DLMs as a non-transformer alternative that decouples generation order from training objective. Operational deployment favors domains tolerant of higher latency for reduced sequential dependency, such as batch code synthesis or offline translation.

Future work must standardize inference budgets before claims of scaling superiority can be tested against transformer baselines on production workloads.