Stop-Think-AutoRegress: Language Modeling with Latent Diffusion Planning

TL;DR

STAR-LDM introduces a novel language model that combines latent diffusion planning with autoregressive generation, enabling global semantic planning and improved performance on understanding benchmarks and controllability.

Contribution

It presents a new language modeling approach integrating diffusion-based planning with autoregression, enhancing global coherence and controllability.

Findings

Outperforms similar-sized models on language understanding benchmarks.

Achieves over 70% win rates in narrative coherence and reasoning evaluations.

Enables attribute control without retraining, with better fluency-control trade-offs.

Abstract

The Stop-Think-AutoRegress Language Diffusion Model (STAR-LDM) integrates latent diffusion planning with autoregressive generation. Unlike conventional autoregressive language models limited to token-by-token decisions, STAR-LDM incorporates a "thinking" phase that pauses generation to refine a semantic plan through diffusion before continuing. This enables global planning in continuous space prior to committing to discrete tokens. Evaluations show STAR-LDM significantly outperforms similar-sized models on language understanding benchmarks and achieves $>70\%$ win rates in LLM-as-judge comparisons for narrative coherence and commonsense reasoning. The architecture also allows straightforward control through lightweight classifiers, enabling fine-grained steering of attributes without model retraining while maintaining better fluency-control trade-offs than specialized approaches.