[arXiv]score: 0.15

FMLM+ Adds Posterior Refinement to Non-Autoregressive Generation

June 24, 2026

FMLM+ extends Flow Map Language Models with masking-style noise schedules, enabling single-step full-sequence generation while scoring per-token global consistency a posteriori. The Posterior Refinement strategy uses these scores for adaptive self-correction, addressing the factorization error that degrades Masked Diffusion Models under parallel decoding. Targets faster inference without the quality collapse of existing non-autoregressive approaches.

HOW THIS AFFECTS YOU

●

builderWorth tracking if you're exploring non-autoregressive decoding for latency-sensitive generation — Posterior Refinement could reduce inference steps without the quality trade-offs of current MDMs.

●

researcherThis bridges MDM flexibility and FMLM quality in a single framework, with a novel inference-time refinement signal worth evaluating against autoregressive baselines on standard benchmarks.

read original ↗arxiv.org

← back to feed