@meihuadang: Introducing Mosaic , a constrained decoding framework for diffusion language models. - significantly improves DLMs' per…
Summary
Mosaic is a constrained decoding framework for diffusion language models that significantly improves performance on function calling benchmarks like JevBench and is compatible with vLLM, SGLang and Hugging Face, supporting models such as DiffusionGemma and LLaDA2.
View Cached Full Text
Cached at: 10/03/26, 12:59 PM
Introducing Mosaic 🧩, a constrained decoding framework for diffusion language models.
- significantly improves DLMs’ performance on general function calling and JevBench.
- compatible with vLLM, SGLang and HF, supporting DiffusionGemma, LLaDA2, and many others.
1/n 🧵
2/n The code is here, with a simple API for vLLM and SGLang. Give it a try!
3/n Autoregressive models enforce constraints by masking out invalid next tokens. But DLMs can generate multiple tokens at once, in arbitrary order. How do we ensure those tokens satisfy the constraints together?
4/n Our key idea: represent finite-automaton constraints as a distribution and multiply it by the DLM’s per-step distribution. This product remains tractable, allowing us to jointly sample tokens that satisfy the constraints.
5/n With corresponding JSON schemas, Mosaic significantly improves function calling accuracy on BFCL.
6/n On JevBench, Mosaic improves DiffusionGemma’s accuracy by constraining its final answer to a JSON object containing a probability for each label: {“probabilities“: {“
Similar Articles
@vllm_project: Congrats to @GoogleDeepMind on DiffusionGemma A 26B diffusion language model on the Gemma4 backbone, and the first dLLM…
vLLM announces native support for Google DeepMind's DiffusionGemma, a 26B discrete diffusion language model that generates 256-token blocks in parallel, enabling low-latency inference at 1200+ tok/s on a single H200.
Multi-Block Diffusion Language Models
This paper proposes Multi-Block Diffusion Language Models (MBD-LMs), extending single-block diffusion to concurrent multi-block decoding with improved training strategies like Multi-block Teacher Forcing and an optimized Block Buffer decoding algorithm. Experiments show increased tokens per forward pass and improved accuracy on benchmarks.
@volokuleshov: New blog post: How to Build a Diffusion Language Model. Diffusion LLMs went from open problem to reality in 2 years (Me…
A comprehensive blog post by Volodymyr Kuleshov's Cornell group explains how to build diffusion language models, covering core techniques like masked diffusion, iterative refinement, variable-length generation, controllable generation, fast samplers, and RL post-training, using open-source models such as Mercury, Gemma Diffusion, and Nemotron Diffusion as examples.
Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM
This paper proposes Dynamic-dLLM, a training-free framework that accelerates diffusion large language models by dynamically allocating cache-update budgets and calibrating decoding thresholds, achieving over 3x speedup on models like LLaDA and Dream while maintaining performance.
ALoDLM: Adaptively Looped Diffusion Language Models
ALoDLM introduces token-adaptive latent recurrence for diffusion language models, allocating extra computation to difficult tokens and using a conditional negative ELBO to jointly learn prediction and compute allocation, outperforming other DLMs and AR baselines at 1.7B and 8B scales across eleven benchmarks.