Last week, the Tencent AI team proposed a novel architecture called "Continuous Autoregressive Language Model (CALM)".

Compared with traditional Transformer, CALM directly reduces the calculation amount used to train large models by 44%, and the calculation amount during inference is also reduced by 34%.

Flow saving: "The core logic of CALM is to compress K consecutive tokens into a semantic vector, and directly predict the "next vector" instead of the "next token" when generating.

Assuming K=4, then the content that originally needed to be generated step by step 4 times can now be completed in just one step, allowing it to output more information in one generation and greatly improve efficiency." Full text address:

Visual summary

Article argument map

Generated from the post's content graph

Mermaid outline

flowchart LR
  format-long_post["longform note"]
  topic-inference["inference"]
  format-long_post -->|covers| topic-inference

Visual structure

Essay structure map

Built from summary and key paragraph positions

Mermaid outline

flowchart LR
  thesis["Last week, the Tencent AI team proposed a novel architecture called \"Continuous Autoregressive Language Mod..."]
  signal["Last week, the Tencent AI team proposed a novel architecture called \"Continuous Autoregressive Language Mod..."]
  operator["Flow saving: \"The core logic of CALM is to compress K consecutive tokens into a semantic vector, and direct..."]
  implication["Assuming K=4, then the content that originally needed to be generated step by step 4 times can now be compl..."]
  thesis -->|frames| signal
  signal -->|develops| operator
  operator -->|lands in| implication

Source: View the original post