Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Stanford and Together AI introduce coarse-to-fine text generation with NCM

AI summary

Stanford Hazy Research and Together AI introduced Next-Concept Modeling (NCM), a text-generation architecture that first creates a coarse representation of an entire passage and then refines it layer by layer. The two-stage system combines a multiscale residual VQ-VAE with a Transformer that predicts all codes in a layer in parallel. The team reported generation speeds 10 to 31 times faster than similarly sized diffusion language models while using four to five times fewer training tokens, and released the models and code under the MIT License.

In-depth insight

NCM offers a distinct alternative to token-by-token autoregressive and diffusion-based generation, but its headline speed and efficiency claims remain unpeer-reviewed and have not been compared with optimized autoregressive systems.

跳脫逐字生成框架,史丹佛新架構 NCM 產字最高快 31 倍

史丹佛大學 Hazy Research 實驗室與 Together AI 10 月 9 日發表論文,提出名為「次概念建模」(Next-Concept Modeling,NCM)文字生成架構。主流大型語言模型一次產生一個詞元(token),NCM 則先產生整段文字的粗略骨架,再逐層補足細節。團隊表示,參數規模相近的條件下,NCM 速度比擴散語言模型快 10~31 倍,訓練所需詞元也少 4~5 倍。

從砌磚到先擬大綱

ChatGPT、Claude、Gemini 等模型多採「自迴歸」(autoregressive)架構,根據前文預測下個詞元,就像工人砌磚,輸出越長步驟越多,寫過的字也無法回頭修改。另一條路線是擴散語言模型,先產生一整段雜亂文字,再像雕刻般反覆修整,但每一步都得處理整段序列,運算成本高。

#Stanford Hazy Research#Together AI#Next-Concept Modeling#VQ-VAE#Transformer#VAR#DUO#FMLM+

Original · 3 min readRead the original
Group chat

No comments yet. Start the conversation.