Mosaic: exact constrained decoding for diffusion LLMs via finite automata, NeurIPS paper

StefanoErmon · x · 2026-10-03

Stefano Ermon's group at Stanford introduced Mosaic, a constrained decoding framework for diffusion language models. Key facts:

• Core idea: treat finite automata as graphical models, enabling exact constrained decoding for diffusion LLMs
• Significantly improves DLM performance on general function calling and JevBench
• Compatible with vLLM, SGLang, and Hugging Face; supports DiffusionGemma, LLaDA2, and others
• To appear at NeurIPS; timely given the rise of diffusion-style LLMs

Original post →

More from Models

• RSIArena, 48 Hours In: Model Catches Test Questions in Its Training Data and Restarts — my_cat_can_code · 2026-10-03
• Spotlight architecture scrutiny: attention-style memory may carry a large constant factor — teortaxesTex · 2026-10-03
• GPT-6.1 Sol reportedly under heavy load; capacity expansion to nearly double serving speed — NandaVegg · 2026-10-03
• One prompt with Opus 5.5 builds a realistic fighter jet game — TAbrodi · 2026-10-03
• Perplexity drops seven open-source projects: SoTA decision model, on-device PII filter, security tools — AravSrinivas · 2026-10-03
• Are hallucinations improving or plateauing? Reddit thread probes AI reliability ceiling — maedhros256 · 2026-10-03

Models channel →