CogGen: A Cognitively Inspired Recursive Framework for Deep Research Report Generation

A research team from Nanjing University and Alibaba Group has published CogGen, a novel AI framework designed to autonomously synthesize in-depth, professional-grade research reports. The system directly addresses the limitations of current linear AI workflows, which often suffer from error accumulation and an inability to globally restructure content based on new insights. CogGen's core innovation is its Hierarchical Recursive Architecture, which mimics human cognitive writing by allowing the AI to recursively plan, draft, and revise its outline and narrative logic, enabling dynamic restructuring throughout the generation process.

To seamlessly integrate visual elements like charts and graphs, the team developed Abstract Visual Representation (AVR), a concise, intent-driven language that describes visual-text layouts. This allows CogGen to iteratively refine multimodal content without the computational overhead of regenerating pixels. The researchers also introduced the Cognitive Load Evaluation Framework (CLEF) and curated a new benchmark from Our World in Data (OWID) to measure performance. Extensive experiments show CogGen achieves state-of-the-art results among open-source systems, generating reports judged to be on par with those written by professional analysts and even surpassing the output quality of Google's Gemini Deep Research tool. The framework's code and dataset have been made publicly available, promising to advance the field of autonomous, high-quality document synthesis.