Fr0zenCrane/UniCoT

Implements a two-level hierarchical reasoning architecture combining macro-level task decomposition (sequential, parallel, and progressive refinement) with micro-level self-reflection mechanisms to enhance multimodal reasoning. Built on the Bagel model foundation, it extends MLLMs to perform interpretable step-by-step reasoning for vision-language tasks including image generation, editing, and visual planning. Training infrastructure supports supervised fine-tuning with evaluation on WISE, KRIS Bench, and RISE Bench datasets.

214 stars.