A breakthrough in artificial intelligence research could revolutionize how language models think. Professor Guojun Qi's team at Westlake University's MAPLE Lab has introduced an innovative reasoning method called the "diffusive divergent chain of thought," specifically designed for diffusion language models.

Traditional AI models typically follow linear thought processes, generating answers through step-by-step reasoning. But human cognition rarely works this way - our minds jump between ideas, make unexpected connections, and explore multiple possibilities simultaneously. The new approach aims to capture this natural complexity.

The diffusive divergent chain allows models to generate intermediate results in any order during reasoning, free from rigid syntactic structures. This flexibility enables exploration of diverse thinking paths, yielding more creative solutions. Early tests show particular promise in mathematical reasoning and code generation tasks.

"Imagine brainstorming with a team where everyone shouts ideas simultaneously rather than taking turns," explains Professor Qi. "That's essentially what we're enabling these models to do - consider multiple possibilities at once rather than following a predetermined path."

The team implemented reinforcement learning to optimize the generation process. Starting from blank sequences, the model progressively builds key information during denoising while using intermediate content to refine final answers. This contrasts sharply with traditional approaches that enforce strict sequential reasoning.

Initial applications in Google's Gemini Diffusion model demonstrate significant potential. The method not only improves reasoning accuracy but also provides valuable insights for future AI training methodologies. Researchers anticipate this could become standard practice for diffusion language models moving forward.

Key Points

1. New "diffusive divergent chain of thought" mimics human non-linear thinking patterns
2. Allows AI models to explore multiple reasoning paths simultaneously
3. Particularly effective for mathematical and coding tasks
4. Uses reinforcement learning to optimize information generation
5. Demonstrated success in Google's Gemini Diffusion model