Salvato in:
Dettagli Bibliografici
Autori principali: Tack, Jihoon, Lanchantin, Jack, Yu, Jane, Cohen, Andrew, Kulikov, Ilia, Lan, Janice, Hao, Shibo, Tian, Yuandong, Weston, Jason, Li, Xian
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2502.08524
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866929712076423168
author Tack, Jihoon
Lanchantin, Jack
Yu, Jane
Cohen, Andrew
Kulikov, Ilia
Lan, Janice
Hao, Shibo
Tian, Yuandong
Weston, Jason
Li, Xian
author_facet Tack, Jihoon
Lanchantin, Jack
Yu, Jane
Cohen, Andrew
Kulikov, Ilia
Lan, Janice
Hao, Shibo
Tian, Yuandong
Weston, Jason
Li, Xian
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2502_08524
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle LLM Pretraining with Continuous Concepts
Tack, Jihoon
Lanchantin, Jack
Yu, Jane
Cohen, Andrew
Kulikov, Ilia
Lan, Janice
Hao, Shibo
Tian, Yuandong
Weston, Jason
Li, Xian
Machine Learning
Computation and Language
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.
title LLM Pretraining with Continuous Concepts
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2502.08524