Human Logical Order Isn't Always Optimal for Model Reasoning

furongh · x · 2026-07-06

A surprising discovery reveals that human logical sequences aren't always optimal for models. Pre-trained diffusion LMs might actually reason best by following a trajectory entirely different from human problem-solving. This points toward a new direction: pursuing "model-aware" thought ordering rather than strictly human-like CoTs.

Related event: ICML 2026 Paper SAS: Optimizing Thought Scheduling in Diffusion LMs(15 posts)→

Original post →

More from Research

Research channel →