NEO原生多模态架构：从零训练挑战视觉编码器 | 论文发布
liuziwei7 · x · 2026-08-16
A new paper titled "From Pixels to Words" proposes principles for constructing native Vision-Language Models (VLMs) and introduces the NEO model family. Addressing the controversy over VLM vision encoders, the authors argue that the best encoder is no encoder and that training from scratch is always superior. NEO models align pixel and word representations within a shared semantic space, integrating separate vision and language strengths. Trained from scratch on 390M image-text pairs, NEO significantly narrows the gap with top-tier modular counterparts.
More from Multimodal
- Open-source anatomy-reveal skill turns Claude into an exploded-view motion design studio — msharmas · 2026-10-16
- Flux 3 Klein vs Flux 2 Klein hands-on: the new base model 'barely usable' for T2I — Both-Rub5248 · 2026-10-16
- USTC paper uses speculative sampling to cut latency in interactive video generation — ustc · 2026-10-16
- 800,000 Gaussian splats run in WebGPU; Casado sees near-infinite scaling — generativist · 2026-10-16
- Nano Banana 2.1 image generation now live in Perplexity Computer for all users — ChrisUniverse · 2026-10-16
- Demo shows precise camera-path control in 3D scenes from a single input image — XRarchitect · 2026-10-16