TBNAI
LatentQuant: Preserving the Policy-Facing Latent Contract under NVFP4 VAE Quantization
Publication: cs.AI updates on arXiv.org
Author: Ziye Deng, Lufang Chen, Shuyu Feng, Zhenwei Duan, Zicong Ye, Yu Sun, Xiaofan Li, Ruyi Gan, Hao Wang, Hao Zhang
Source published: 06/10/2026, 04:00:00 UTC
Briefing created: 06/10/2026, 05:18:13 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 04:18:24 UTC.
Briefings
This article presents LatentQuant, a two-stage NVFP4 quantization-aware training framework designed to preserve the policy-facing latent contract in video action models. The study focuses on Wan2.1 and Wan2.2, demonstrating that joint quantization-aware training can recover high reconstruction fidelity and near-baseline control performance. Key findings include a 1.17x-1.26x end-to-end VAE speedup on NVIDIA B300 GPUs compared to BF16 cuDNN, while LIBERO success rates improved from 10.5% to 95.75% and RoboTwin success rates reached 68.8%.
Key points
- LatentQuant is a two-stage NVFP4 framework that first aligns the quantized encoder with its high-precision counterpart before freezing it while adapting the decoder.
- On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17x-1.26x end-to-end VAE speedups over BF16 cuDNN.
Context and limitations
The study covers only Wan2.1 and Wan2.2; LIBERO success rates collapsed from 95.5% to 10.5% in controlled decoder-only experiments, and persistent latent drift remains an observed phenomenon.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 06/10/2026, 05:18:13 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeid5ajluy3hcm4gnrlbgqzu6etqowh7b4tapurujxk3zanx5ioitwu