CoG adds cognitive cycles to GraphRAG

CoG’s strongest idea is treating retrieval as adaptive investigation rather than a one-shot lookup, though its infrastructure demands make experimentation far easier than production deployment.

• –Text-derived entities can repair gaps in incomplete knowledge graphs and redirect exploration toward useful evidence
• –The framework improves smaller backbones substantially, with CoG-8B outperforming several direct-prompting larger-model baselines
• –Reported efficiency gains are meaningful: 2.6 documents per question versus 32.5 for ToG-2 on average
• –Iterative planning and reflection add latency, token cost, and dependence on an instruction-following backbone
• –Running the full Wikidata stack requires roughly 500GB of storage and 500GB of RAM, a major barrier for individual developers