LaDiR: Latent Diffusion Strengthening LLM Text Reasoning

LaDiR (Latent Diffusion Reasoner) is a framework that reconstructs chain-of-thought (CoT) reasoning in LLMs using a VAE and a latent diffusion model. It targets the problem that existing autoregressive decoding cannot sufficiently look back at previous tokens, limiting exploration and revision.

First, it creates a latent reasoning space that compresses text reasoning steps into thought token blocks, securing a more compact representation while preserving semantic information and interpretability.

Then, latent diffusion applying a blockwise bidirectional attention mask progressively refines these blocks. This structure enables planning and revision over a longer range, and leverages adaptive test-time compute to generate multiple reasoning paths in parallel.

• Evaluation: Verified across multiple tasks including math reasoning and planning benchmarks.
• Results: Consistently improved accuracy, diversity, and interpretability compared to existing autoregressive, diffusion-based, and latent reasoning methods.

LaDiR presents a new paradigm for planning more flexibly, exploring more broadly, and refining text reasoning better through latent diffusion.