SONAR-LLM: Autoregressive Transformer that Thinks in Sentence Embeddings
  and Speaks in Tokens

Nikita Dragunov; Temurbek Rahmatullaev; Elizaveta Goncharova; Andrey Kuznetsov; Anton Razzhigaev

返回文献探索

Paper · arXiv 2508.05305

SONAR-LLM: Autoregressive Transformer that Thinks in Sentence Embeddings and Speaks in Tokens

Nikita Dragunov, Temurbek Rahmatullaev, Elizaveta Goncharova, Andrey Kuznetsov, Anton Razzhigaev

49 upvotesAugust 7, 2025arXiv 预印本

AI 摘要

SONAR-LLM, a decoder-only transformer using token-level cross-entropy in the SONAR embedding space, achieves competitive text generation quality without diffusion sampling.

Large Concept ModelLCMSONAR-LLMdecoder-only transformerSONAR embedding spacetoken-level cross-entropydiffusion objectiveslikelihood-based training signal

Abstract

The recently proposed Large Concept Model (LCM) generates text by predicting a sequence of sentence-level embeddings and training with either mean-squared error or diffusion objectives. We present SONAR-LLM, a decoder-only transformer that "thinks" in the same continuous SONAR embedding space, yet is supervised through token-level cross-entropy propagated via the frozen SONAR decoder. This hybrid objective retains the semantic abstraction of LCM while eliminating its diffusion sampler and restoring a likelihood-based training signal. Across model sizes from 39M to 1.3B parameters, SONAR-LLM attains competitive generation quality. We report scaling trends, ablations, benchmark results, and release the complete training code and all pretrained checkpoints to foster reproducibility and future research.