When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues
Factual evidence
What the source reports
Researchers identify 'State Integrity Collapse' in LLMs, where models retain explicit instructions in memory but fail to apply them.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 23 October 2026
- Collected by OneBench
- 24 Oct 2026, 03:01 UK
Stored source excerpt
arXiv:2610.31314v1 Announce Type: new Abstract: Large language models can repeat an instruction while failing to keep obeying it. We study this recall-compliance dissociation in long-horizon…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
Models managing long-horizon conversations may silently violate risk constraints despite successfully recalling system prompts during verification.
Do what
Review conversational model evaluation frameworks to ensure compliance testing covers multi-turn state integrity rather than single-turn prompt recall.