Thinking into the Future: Latent Lookahead Training for Language Models

ICLR 2026 Conference Submission19610 Authors

Keywords: latent reasoning, language models

Abstract:

Autoregressive language models trained with next-token prediction generate text by sampling one discrete token at a time. This forces the model to commit early, preventing exploration of multiple plausible continuations. Furthermore, each token is predicted in a single forward pass, which might limit the model’s expressiveness in cases where difficult tokens require inherently more compute. Towards this end, we introduce latent lookahead, a training strategy that enables models to think before answering: at selected positions in the sequence, before committing to the next token, the model performs a multi-step lookahead in latent space. Instead of sampling future tokens, we leverage the network’s latent space by recursively feeding its hidden states back into the context for τ steps, investing more compute on predicting that token. This produces τ latent predictions that are supervised against the next τ ground-truth tokens, encouraging the model to “look
ahead”. We show that latent lookahead substantially outperforms autoregressive baselines on planning tasks such as maze solving, Sudoku, and ProsQA, where foresight is essential. Finally, we demonstrate how to endow pretrained models with this ability during supervised fine-tuning and evaluate the resulting models on standard reasoning benchmarks.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 19610