LLM Pretraining with Continuous Concepts

TL;DR

CoCoMix introduces a novel pretraining framework that combines discrete token prediction with continuous concepts, improving efficiency, performance, and interpretability of large language models across various tasks.

Contribution

It proposes a new pretraining method that integrates continuous concepts learned from autoencoders with token prediction, enhancing model efficiency and interpretability.

Findings

Outperforms standard next token prediction on multiple benchmarks.

Increases sample efficiency in language modeling and reasoning tasks.

Enhances interpretability and steerability of language models.

Abstract

Next token prediction has been the standard training objective used in large language model pretraining. Representations are learned as a result of optimizing for token-level perplexity. We propose Continuous Concept Mixing (CoCoMix), a novel pretraining framework that combines discrete next token prediction with continuous concepts. Specifically, CoCoMix predicts continuous concepts learned from a pretrained sparse autoencoder and mixes them into the model's hidden state by interleaving with token hidden representations. Through experiments on multiple benchmarks, including language modeling and downstream reasoning tasks, we show that CoCoMix is more sample efficient and consistently outperforms standard next token prediction, knowledge distillation and inserting pause tokens. We find that combining both concept learning and interleaving in an end-to-end framework is critical to performance gains. Furthermore, CoCoMix enhances interpretability and steerability by allowing direct inspection and modification of the predicted concept, offering a transparent way to guide the model's internal reasoning process.