First-hand information for everyone
Main discovery Researchers identified a failure mode they call State Integrity Collapse (SIC). SIC occurs when a model can restate or recall a persistent constraint on request, but then fails to apply that constraint in its next generated reply. In other words, the model’s declarative memory and its generative behavior become decoupled as dialogue state grows. (SIC defined: F1)
How the study tested this The team built StateTrack‑Bench, a controlled suite of multi-turn conversations. Each instance injects one persistent constraint, then runs distractor turns, and finally probes the model at later depths. The benchmark covers five constraint categories: numerical, format, role/persona, factual‑memory bindings, and logical rules. Dialogues use three phases (setup, distractor, test) and are probed at turn depths T = {3, 5, 7, 10, 12}. (F2, F4)
Each model was run on 125 scenario–depth probes (25 per category); deterministic decoding and fixed prompts and random seeds were used for reproducibility. (F3, F7)
Metrics and audit The headline metric is strict joint retention (R_m). A response counts only if it both completes the immediate task and satisfies the persistent constraint. To separate forgetting from application failures, the authors run a forked recall/compliance audit. For each probed history they (a) ask the model to perform the task (compliance branch) and (b) ask the model to restate the constraint (recall branch). This yields instance‑level measures of SIC and its signed components. (F5, F6)
Concrete results Across the seven snapshots and the five probe depths, strict joint retention varied a lot by model. Observed overall scores ranged from 62.4% (Gemini‑3.1‑Pro) to 100.0% (DeepSeek‑v4‑Flash). Degradation was concentrated in symbolic constraints. Factual‑memory and role/persona bindings were retained at about 95% on average, while numerical constraints averaged ≈87.7%, format ≈85.5%, and logical‑rule ≈83.5%. (F8, F13)
The forked audit showed most failures were integrity failures (SIC+): the model could recall the constraint but failed to follow it in the next answer. Across the full test set 88.6% of observed failures were SIC+; excluding the brittle outlier Gemini‑3.1‑Pro this remained 84.4%. (F9)
Extended pilots to greater depth confirmed the dissociation. In a depth‑50 numerical analysis recall stayed at 100.0% while compliance dropped to 79.2%, producing SIC+ = 20.8% for those probes. Integrity failures dominated at long depths as well. (F11)
A simple inference‑time mitigation, Structured State Anchoring (SSA), helped in stress tests. SSA—which prepends a structured active‑state and uses verification with retries—recovered up to +24.4 percentage points on immutable constraints for Gemini‑3.1‑Pro in the reported stress tests. (F10)
Robustness checks The observed patterns and model ordering were robust to decoding temperature. Re‑evaluation at temperatures up to 0.7 shifted scores by at most about 2.1 percentage points and preserved the model ranking. The primary automated judge was GPT‑5.4 at temperature 0 using a fixed judge prompt. Judge cross‑validation with architecturally distinct judges reported high inter‑judge agreement. (F13, F12)
Significance and supported caveats The results show that single‑turn instruction‑following scores do not guarantee a model will keep applying remembered constraints as dialogue state grows. The dominant failure mode among observed collapses is behavioral drift: the rule is present in memory but no longer governs generation. This distinction matters for mitigation: retrieval or longer context can help when state is inaccessible, but control‑ or reranking‑style interventions (or inference‑time fixes like SSA) are needed when recalled rules are not enforced. (F1, F9, F10)
Limitations reported by the authors The core benchmark uses 25 probes per category, which limits fine‑grained within‑category statistical power. The evaluation relies on LLM‑based judges, although cross‑validation showed high agreement. Distractor turns were intentionally cooperative rather than adversarial. The benchmark tests only one persistent constraint per dialogue, so multi‑constraint interference and dynamic revocation remain unmeasured. Full evaluation beyond 50 turns and larger, cross‑lingual studies were left for future work. (F14)