FLM Does One-Step Language Modeling via Continuous Denoising, 8.3x Faster, NeurIPS 2026
ArashVahdat · x · 2026-09-25
A KAIST/CMU paper accepted at NeurIPS 2026 challenges the assumption that discrete diffusion is the right path for fast text generation. FLM uses flow-based continuous denoising on one-hot tokens, outperforming SOTA discrete diffusion models in many-step regimes; its distilled FMLM achieves few-step parallel generation that matches 8-step baseline quality in a single step, with an 8.3x speedup on LM1B and OpenWebText.
More from Research
- Study: Reasoning hurts 15.7% of multimodal embeddings; training-free SURE router fixes it — _reachsumit · 2026-09-25
- Google scales learned cross-task relationships in YouTube's production recommender — _reachsumit · 2026-09-25
- ByteDance's OneTrans-V2 unifies retrieval, pre-rank and fine-rank, lifting GMV 9.74% — _reachsumit · 2026-09-25
- Training-free Seek framework beats BM25 by 82% on BRIGHT via self-evaluative iteration — _reachsumit · 2026-09-25
- ByteDance's X-Rec does generative retrieval via flow matching with 3.46x throughput — _reachsumit · 2026-09-25
- Multi-agent AI teams are swayed by deceptive agents even in the minority, paper finds — sethlazar · 2026-09-25