Ai2's StateBench Drops Frontier Models to 13% on State Tracking

The Allen Institute for AI released StateBench, a benchmark revealing that current frontier language models like Gemini 3.8, Claude Opus 5.5, and GPT-6 variants collapse to 13% accuracy or lower on long, complex state-tracking tasks. The research highlights that transformer architectures cannot generalize these tasks well, unlike Gated DeltaNet and hybrid recurrent models which generalize almost perfectly on extended sequences.

This Wire brief sits within Fusion42's coverage of AI Frontier Models and AI Infrastructure.