Stop-Think-AutoRegress: Language Modeling with Latent Diffusion Planning
This addresses the challenge of enabling global planning in language models for better coherence and reasoning, representing a novel method rather than an incremental improvement.
The paper tackles the problem of limited global planning in autoregressive language models by introducing a 'thinking' phase that refines semantic plans through diffusion before token generation, resulting in significant performance improvements on language understanding benchmarks and over 70% win rates in LLM-as-judge comparisons for narrative coherence and commonsense reasoning.
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.