LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning
10 hours ago
- #natural language processing
- #reasoning models
- #machine learning
- LaDiR is a novel reasoning framework combining latent diffusion models with LLMs to enhance text reasoning by creating a structured latent reasoning space using a VAE.
- It encodes text reasoning steps into blocks of thought tokens, offering compact and interpretable representations, and uses a latent diffusion model for iterative refinement with adaptive test-time compute.
- The approach enables parallel generation of diverse reasoning trajectories, improving holistic planning and revision over autoregressive methods.
- Empirical evaluations on mathematical reasoning and planning benchmarks show LaDiR improves accuracy, diversity, and interpretability compared to existing methods.
- The paper introduces a new paradigm for text reasoning with latent diffusion and was presented at a workshop on latent and implicit thinking at ICLR 2026.