Apple researchers built an AI that tests several ideas in parallel before answering

Apple and UC San Diego researchers introduced LaDiR, a framework that uses latent diffusion to explore multiple hidden reasoning paths before generating the final answer autoregressively.

In tests on math, coding, and planning benchmarks, LaDiR outperformed baseline approaches, with especially strong gains on harder tasks.

The method was evaluated with models like LLaMA 3.1 8B and Qwen3-8B-Base, showing better reasoning diversity without changing the base model.