When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues
StateTrack-Bench studies whether LLMs can keep obeying a persistent rule across multi-turn conversations. The authors define State Integrity Collapse (SIC): a model remembers a constraint yet fails to enforce it in later answers, a failure mode they distinguish from ordinary forgetting. They evaluate seven pinned model snapshots across five constraint categories and five dialogue depths, showing that compliance decay is concentrated in symbolic rules and in one notably brittle model, and they argue convincingly that single-turn instruction-following metrics are insufficient for deployed assistants.
The paper makes a sharp conceptual and methodological contribution by isolating recall–compliance dissociation via a forked audit protocol. Its empirical base is more tentative: only 125 probes per model and 25 per category limit effect estimates, and the headline aggregate is heavily skewed by one brittle model. The authors correctly treat results as diagnostic signals rather than robust rankings, but readers should note that the striking 88.6% integrity-failure headline is severity-skewed by Gemini-3.1-Pro and that excluding this outlier leaves most frontier models near ceiling.
The forked SIC audit is the paper's strongest element: by splitting each dialogue prefix into independent recall and compliance branches, the authors cleanly separate memory loss from state-application failure. The signed metrics $\operatorname{SIC}^{+}$ and $\operatorname{SIC}^{-}$ prevent opposite errors from cancelling, and the five-category taxonomy yields an interpretable pattern in which symbolic constraints—numerical, format, and logical—are systematically more brittle than semantic ones like factual memory and role persona.
The primary limitation is sample size. With only 25 scenarios per category, the benchmark cannot support fine-grained model comparisons or reliable within-category failure rates; while the authors appropriately abstain from significance-ranked leaderboards, the small n still bounds inferential confidence.
A second concern is outlier influence: Gemini-3.1-Pro is so far below the other six models that its failures dominate aggregate statistics, and the headline integrity-failure share drops sharply when it is excluded. Third, the evaluation uses single-constraint dialogues with cooperative distractors, so it misses multi-constraint interference, user-initiated revocation, and adversarial override that are common in real deployment.
The authors fairly distinguish their setting from short-horizon instruction following and explicit slot-tracking benchmarks, and they acknowledge that prior long-context work tests retrieval rather than behavioral compliance. Multi-judge cross-validation is strong, with $\kappa = 0.89$, but the extended-depth and cross-lingual pilots remain small scope checks rather than robust generalizations. The link to Liu et al.'s lost-in-the-middle effect is apt, because the forked audit shows that even when recall succeeds, compliance can fail—a genuine extension of the position-bias literature.
Reproducibility is aided by fixed seeds, deterministic decoding, pinned provider snapshot identifiers, and public release of prompts, logs, and evaluation scripts. The exact endpoint strings are stored as reproducibility handles. The main barriers to independent replication are access to the same historical model snapshots—some of which may be unreleased provider versions—and the reliance on a single primary LLM judge whose own weights may be inaccessible or unstable.
Large language models can repeat an instruction while failing to keep obeying it. We study this recall-compliance dissociation in long-horizon dialogue and define State Integrity Collapse: a model preserves a stated constraint in memory but stops applying it to later turns. We introduce StateTrack-Bench, a controlled suite of conversations covering numerical, format, role/persona, factual-memory, and logical-rule constraints. Across seven provider-pinned model snapshots and five dialogue-depth probes, strict joint retention ranges from 62.4% to 100.0%, with degradation concentrated in symbolic constraints and in one especially brittle model. A forked recall audit shows that 88.6% of observed failures are behavioral-drift failures rather than ordinary forgetting; excluding the brittle outlier preserves this diagnosis (84.4% of failures). Extended evaluations at depths up to 50 turns confirm that the dissociation persists beyond the core benchmark, while Structured State Anchoring recovers up to 24.4 points on immutable constraints. The results show that reliability cannot be inferred from single-turn instruction following alone; evaluation must track whether remembered constraints remain operational as dialogue state grows. Code and artifacts are available at: https://github.com/xzzvsxd/StateTrack.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.