StateBench tests long-context state tracking in LLMs

Researchers introduced StateBench, a benchmark measuring how well language models track evolving entity states in long documents, across tasks of varying difficulty. They found recurrent and hybrid architectures generalize much farther than transformers when trained on these tasks, and frontier models such as Gemini 3.8 Flash and Claude Opus 5.5 were among those evaluated.

• Preprint probes how source attribution shifts LLM argument ratings2026-11-04
• User has Claude hand-edit Gemma neurons to tune Italian translation2026-11-03
• Claude Opus 5.5 Can Draw Permit-Level Architectural Drawings2026-11-02
• Anthropic adds Claude rule against recreating existing games2026-11-02
• Minecraft test exposes AI swarm failures2026-11-02