NVIDIA's Nemotron-Labs-Diffusion-Uni built a unified discrete diffusion model that beat the prior unified diffusion state of the art by up to 48.8 points on multimodal reasoning while running 552 tokens per second
AspireAI community · arXiv (NVIDIA Labs)
Nemotron-Labs-Diffusion-Uni (NLD-Uni), an arXiv paper submitted October 27, 2026, unifies visual understanding and visual generation in a single discrete diffusion language model using two ingredients: a cross-paradigm on-policy distillation framework that transfers an autoregressive VLM's understanding and a continuous flow model's generation ability into one student, and a progressive self-refining diffusion process that corrupts each token across multiple training steps instead of masking it in one shot. On the authors' multimodal reasoning table NLD-Uni scores 60.7 / 70.5 / 89.9 / 94.2 / 84.0 / 76.2 / 85.0 / 94.1 / 95.4 / 49.2 / 62.1 / 78.3 against the previous unified diffusion state of the art, with per-benchmark deltas from +0.1 up to +48.8 points. The paper reports NLD-Uni surpassing prior discrete diffusion models on text-to-image generation and only underperforming GPT-4o and the Qwen-Image teacher. Its vanilla sampler runs at 552.58 tokens per second against 202.74 for an autoregressive baseline, and relaxing the confidence threshold reaches 997.95 tokens per second with minor degradation. The honest limitation is scope: the authors state the model lacks fine-grained text rendering because such data was not widely present in the training mix, and no public weights or code repository is linked from the paper.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.