NCM brings VAR-style multiscale discrete generation to text with parallel per-level decoding
burny_tech · x · 2026-10-23
Hermann Kumbong outlines NCM (Next-Concept Model), a two-stage approach that ports the VAR idea from image generation to text:
- Learn: a multiscale VQ-VAE encodes text into a hierarchy of discrete residual codes, from coarse concepts to fine details. The author notes VAR already does this for images, but making it work for text required substantial modeling changes.
- Generate: a transformer predicts the hierarchy level by level, generating all codes within each level in parallel, conditioned on the cumulative reconstruction of earlier levels, with block-diagonal attention to limit cost.
The accompanying diagram details the tokenizer's encode/decode pipeline and the masked attention structure for level-wise generation.
More from Research
- ruff author asks: how hard is scaling AI chip design from RISC-V to x86? — charliermarsh · 2026-10-23
- Transformer jointly models electrons and magnetism with 0.01% energy error — bravo_abad · 2026-10-23
- Google Cloud AI releases PLE benchmark; all frontier models score under 50% — sainingxie · 2026-10-23
- AI x Bio's wild 48 hours: OpenAI buys Orchestra Bio, 15+ launches — vaibhavbetter · 2026-10-23
- PL researchers unusually upbeat about Anthropic–Theorem verification work — xennygrimmato_ · 2026-10-23
- CRYSTALMINER: an OSI hackathon project teaching crystallography via 23 Minecraft-like symmetry worlds with an AI tutor — CatAstro_Piyush · 2026-10-23