AI Briefing#diffusion-language-models
The latest AI and developer news about #diffusion-language-models, with the original source and a short summary.
The latest AI and developer news about #diffusion-language-models, with the original source and a short summary.
NCM uses a two-stage coarse-to-fine framework to generate text in parallel, outperforming diffusion language models in speed and perplexity on standard benchmarks.
The model features a 35B-parameter mixture-of-experts backbone with approximately 3B active parameters per token and a 128K context window.
By generating guidance signals from the internal states of existing models without additional inference passes, it elevates unconditional generation performance to SOTA levels.
Diffusion language models enable parallel generation but showed limitations in long contexts and batch processing.
FuriosaAI presented four papers at ICML 2026 addressing efficiency in LLM inference, training, and serving.
LLaDA2.2 has been released, overcoming the fixed-length output limitation of diffusion language models and improving agentic performance.
Ant Group has released the LLaDA2.0 series, a large-scale diffusion language model with a 100B-parameter MoE architecture.
inclusionAI released LLaDA2.2-flash, an agent-oriented diffusion language model supporting 128K context and Levenshtein Editing.
The paper proposes MBD-LM, a Multi-Block approach designed to improve the parallel decoding performance of Block Diffusion Language Models (BD-LMs).
By recycling information from discarded tokens through RCD technology, the accuracy and inference speed of diffusion language models have been dramatically improved.
This work proposes a new policy learning method that maximizes the efficiency of token selection (unmasking) in Diffusion LLMs through reinforcement learning.
ByteDance has released checkpoints for Cola-DLM, a continuous latent diffusion language model.
A diffusion LLM has achieved both AR-level quality and parallel generation speed together.
I-DLM simultaneously boosts DLM quality and speed through introspective consistency.