Apple and UC San Diego Team Propose LaDiR Framework to Enhance LLM Reasoning Capabilities

Apple Inc. and the University of California, San Diego team have proposed the LaDiR framework to enhance the text reasoning capabilities of large language models (LLMs). This framework is not a new model but a general one that can be superimposed on existing models, optimizing output quality by changing how models think.

LaDiR combines diffusion and autoregressive generation paradigms: during reasoning, it uses diffusion models to explore multiple paths in parallel—each starting from random noise and gradually optimizing into coherent reasoning steps; the final output stage uses autoregressive models to ensure coherence. The framework also introduces a diversity encouragement mechanism to avoid premature convergence of paths and generate diverse candidate answers.

The research team tested LaDiR on LLaMA3.1 8B and Qwen3-8B-Base. It achieved higher accuracy in math benchmarks and outstanding performance on out-of-distribution tasks; in HumanEval code generation, it produced more reliable results, with difficult problems outperforming standard fine-tuning methods. For puzzle planning tasks, it explored a wider solution space, but its single-attempt accuracy was slightly lower than specialized models optimized for specific tasks.