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.
@article{arxiv.2508.05305,
title = {SONAR-LLM: Autoregressive Transformer that Thinks in Sentence Embeddings and Speaks in Tokens},
author = {Nikita Dragunov and Temurbek Rahmatullaev and Elizaveta Goncharova and Nikita Kurdiukov and Aysel Mirzoeva and Anna Borisiuk and Andrey Kuznetsov and Anton Razzhigaev},
journal= {arXiv preprint arXiv:2508.05305},
year = {2026}
}