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

TL;DR

SONAR-LLM is a decoder-only transformer that generates text by thinking in sentence embeddings, combining the semantic abstraction of LCM with likelihood-based training, achieving competitive quality across various sizes.

Contribution

It introduces SONAR-LLM, a hybrid model that unifies sentence embedding thinking with token-level training, improving upon previous Large Concept Models.

Findings

Achieves competitive generation quality across model sizes from 39M to 1.3B parameters.

Retains semantic abstraction while using likelihood-based training.

Provides comprehensive benchmarks and reproducibility resources.

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.