MUSE-AI-Lab/StateTrack

Yancheng Zhu · Yifan Yuan† School of Artificial Intelligence, Shenzhen University †Corresponding author EMNLP 2026 (Main Conference) StateTrack-Bench evaluation protocol: a persistent constraint is planted in the dialogue, probed at increasing depths, and every failure is re-examined with a forked recall test. TL;DR: An assistant can repeat a user's rule and still stop obeying it. We call this recall–compliance dissociation State Integrity Collapse (SIC): the model keeps the constraint in memory but no longer applies it to later turns — unlike ordinary forgetting, longer context or memory retrieval does not fix it. StateTrack-Bench plants one persistent constraint (numerical, format, role/persona, factual-memory or logical-rule) in a conversation, checks compliance at increasing dialogue depths, and uses a forked recall audit to separate forgetting from remembering-without-complying. Across seven provider-pinned models, 88.6% of failures are behavioral drift rather than forgetting, and Structured State Anchoring (SSA) recovers up to 24.4 points. All reported tables are…