LLMs retain rules but stop following them in long conversations
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a critical failure mode in long-context LLM conversations where models retain stated constraints in memory yet stop enforcing them over multiple turns. The new StateTrack-Bench evaluation reveals that compliance retention ranges from 62% to 100% across model snapshots, with failures concentrated in symbolic constraints rather than genuine forgetting. This behavioral drift, affecting nearly 89% of observed failures, exposes a gap between what models claim to remember and what they actually apply, raising questions about reliability in extended interactions where consistency matters most.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The critical finding isn't that models fail on long conversations, but that they retain stated constraints in their weights yet stop applying them mid-dialogue. This points to an enforcement problem, not a memory problem, which changes how teams should debug and mitigate the issue.
This connects directly to the OptionWorld benchmark from October 20th, which also identified context retention failures across multi-turn tasks. But StateTrack-Bench adds precision: it isolates whether the model actually forgot a constraint or simply stopped enforcing it. The same pattern appears in the KEEP-Bench work from October 19th, where models produce correct outputs despite losing evidence in their reasoning chain. Together, these three papers suggest a systematic gap between what models claim to track and what they actually apply, affecting both long-horizon planning and constraint compliance.
If StateTrack-Bench results hold when tested on models released after October 2026, and if the 89% behavioral-drift figure remains consistent across different constraint types (not just symbolic ones), that confirms this is an architectural issue rather than a training artifact. Watch whether any major model provider publishes mitigation results on this specific benchmark within six months.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
OptionWorld benchmark exposes LLM failures in multi-day task execution
Researchers have built OptionWorld, a diagnostic benchmark exposing critical gaps in how LLMs handle multi-day personal assistance tasks. The work reveals that current models fail to retain key facts across conversation turns, miss actionable opportunities, and apply unnecessary constraints to problems. These findings matter because production AI assistants are already deployed in long-horizon workflows, yet…
MentionsStateTrack-Bench
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.