Hugging Face Releases StateBench Benchmark for Long-Context State

Hugging Face released StateBench, an Allen AI benchmark testing language models' ability to track entity state changes across long contexts. It defines four difficulty tiers based on assignment types: immutable (r-trivial), value reassignment (aperiodic), self-referential updates (soluble), and reference-based dependencies (general finite). Standard transformers and frontier models handle easier tiers but collapse on general finite tasks. Hybrid and Gated DeltaNet architectures general