MindLoom: Framework for Frontier-Level Reasoning Data Synthesis

A new framework called MindLoom proposes synthesizing frontier-level reasoning data for large language models through compositional thought mode engineering. The approach views problem difficulty as arising from the accumulation of atomic knowledge-reasoning transformations, termed thought modes. MindLoom first decomposes hard problems with verified solutions into thought mode chains, then trains a retrieval model to match problem states with compatible thought modes. This aims to address limitations in existing synthesis methods, such as narrow diversity and unstable difficulty control. The research is published on arXiv under identifier 2605.21630.