When State Changes Hide in Noun Phrases

Long stories, multi-session conversations and evolving news all report change. Systems that read them must notice when a fact stops being true and must keep track of what holds now. Some changes are announced with a verb: "Jane Cooke was appointed director." Many arrive inside a noun phrase instead: "interim director Jane Cooke," "the former prime minister," "Smith's resignation last spring." Information-extraction research has long documented that extractors see verb-mediated facts far more reliably than these nominal ones. Separately, memory and editing research has shown that language models keep asserting outdated facts across long contexts. Nobody has asked whether the second failure is partly caused by the first. This note proposes to measure that directly. We build long, coherent narratives in which the same timeline of state changes is written once with verbs and once with only noun phrases, verified by humans in small segments. We then ask how much of "state drift" is really a reading problem about linguistic form rather than a reasoning problem about time.

The stakes are practical. If memory systems lose a large share of accuracy whenever an update is phrased nominally, then today's state-tracking benchmarks, which mostly phrase updates as explicit revisions, overstate how well these systems work on real text. The repair would also be cheap: a form-aware ingestion step, not a new reasoning architecture. If no such gap exists, that is worth knowing too. It would mean modern models have absorbed nominal semantics, and the field should stop attributing drift to extraction.

Core Claim

Our hypothesis is that, holding the underlying timeline fixed, updates expressed only through nominal forms produce substantially more stale answers than the same updates expressed verbally. These forms include nominalizations, appositive or pre-nominal titles, temporal modifiers such as "former," "acting" or "ex-," and compound nouns. We expect the gap to persist even when the relevant text is guaranteed to be in context.

We further predict that one class of nominal form is especially harmful. These are phrases that announce a change while lexically repeating the old state, such as "former CEO Smith." They should amplify the "salience trap" identified in Can Agent Memory Systems Track Evolving State?, because each mention re-exposes the model to the superseded attribute. The same class should also be where the "unknown state" failure from Event-level Knowledge Editing concentrates. "The former captain" invalidates a fact without naming a successor. That is exactly the case where ELKEN found models hallucinate a replacement instead of representing a vacancy.

The closest existing work differs in its controls. StateMem and its benchmark (Can Agent Memory Systems Track Evolving State?) motivate this proposal: they show drift dominates errors even under oracle retrieval, but their scenarios phrase updates as explicit revisions. PARALLELEVENTS (Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs) builds coherent synthetic future worlds, but its graphs are built from action events. Its reported "numerical gap" on exact non-verbal properties is suggestive evidence for our hypothesis, though not a test of it. SUMIE (SUMIE: A Synthetic Benchmark for Incremental Entity Summarization) tracks evolving nominal attribute–value pairs, but only in isolated paragraphs and without controlling surface form. No existing work varies linguistic form while holding content and discourse fixed. That manipulation is the new ingredient here.

Research Program

Building the matched-form benchmark. We start from programmatically generated state timelines. Each timeline is a set of entities with roles, affiliations, locations and quantities that change over time. The changes include explicit vacancies, where a role is vacated with no successor, and dependent attributes, where for example a person's office location follows from their role. Program-guided generation of this kind follows the template-and-execute approach of ToolQA: A Dataset for LLM Question Answering with External Tools and the program-guided adversarial construction in ExploreToM. Because gold answers to every "what holds now" question come from the program, they are never written by a model.

We then render each timeline as a long narrative, using a recursive plan-then-draft procedure in the style of Re3: Generating Longer Stories With Recursive Reprompting and Revision and the seed-to-subplan pipeline of Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs. Re3 tracks character attributes in a dictionary but treats them as static. We use its machinery only to keep the prose coherent, not as the object of study.

The key step is a set of five sibling renderings of the same narrative skeleton. They are identical except for how each update sentence is phrased: as a verbal clause, a nominalization, an appositive or pre-nominal title, a temporal-modifier phrase, or a compound noun. Each update slot also gets distractor mentions that repeat the old state, so salience is controlled across conditions. We target about 150 timelines, which gives 750 narratives. Lengths are stratified from roughly 20k to 200k tokens, with current-state probes after each update and vacancy probes after each invalidation.

Human verification in segments. SUMIE's human audit found that unguided synthesis of evolving entity tables produced 45% redundant values and 25% unsupported attribute–value pairs. The SNaC study (SNaC: Coherence Error Detection for Narrative Summarization) found holistic human judgment of long narratives nearly random (Krippendorff's α of 0.19), rising to 0.69 when annotators judge short segments in rolling context. We therefore never ask humans to approve whole narratives.

Annotators see a 2–4 sentence window around each update slot, plus the program's intended update. They answer three narrow questions: does the window entail the intended change, does it entail anything else, and is the nominal rendering felicitous? The last question matters because "former" carries implicatures a string check cannot verify. Failed slots are regenerated. This author–reviewer division of labor is what Building a Role Specified Open-Domain Dialogue System Leveraging Large-Scale Language Models found roughly 13 times more efficient than authoring from scratch. We will report our own cost per verified slot and our own agreement rather than assume theirs.

Form-aware state ingestion. As a candidate fix, we add a front end to a structured state memory. It detects nominal state cues and rewrites each one into an explicit, dated update statement before the memory write. The update vocabulary includes a first-class "vacant/unknown" value.

The detector combines two sources. One is an attribute ontology mined in the manner of ReNoun: Fact Extraction for Nominal Attributes, which found that knowledge-base attributes like CEO or spouse are inherently nominal. The other is a small closed class of temporal modifiers, and it is trained on a held-out split of our verified slots. We place this front end before StateMem's dependency-graph rechecker. Our contribution is only the ingestion step, so any gain can be attributed to it.

Generation. To connect drift to global narrative coherence, we continue each narrative with a Re3-style drafter that must write the next chapter consistently with the current state. We count state-inconsistency errors with SNaC's InconE and CharE categories, scored both by humans and by SNaC's trained detector.

Baselines

The naive reference point is a long-context model given the entire narrative with no memory structure. A second reference point is the same model under oracle retrieval, where the update window and the distractors are guaranteed in context. This separates reading failures from retrieval failures, which is the separation StateMem used to show that drift survives perfect retrieval.

The strongest existing state tracker is StateMem with StateMemWrapper, from Can Agent Memory Systems Track Evolving State?. That paper reports current-state accuracy rising from 0.205 to 0.363 on the DeepSeek-V4-Flash backbone, and wrapper lifts of +32 to +67 points across six memory backends on StateMemBench. It is the right competitor because it is designed for exactly the drift we study and already handles dependencies. We run it with its released prompting on the same backbone, the same narratives and the same token budget as our method. We also give it the same number of development-set tuning passes for its parsing prompt that our detector receives.

A second strong baseline is the LIGHT memory from Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs. LIGHT reports gains of 3.5% to 12.69% over the strongest baselines on BEAM, where contradiction resolution was the weakest category; we found no reported number for that category on Lacuna. LIGHT represents the compressed-scratchpad approach.

The ablation that isolates our ingredient is StateMem whose parsing prompt has been explicitly instructed, with few-shot examples, to extract nominal and appositive updates and vacancies. If this prompt-only variant closes the form gap as well as the trained front end does, then our contribution reduces to a benchmark finding plus a prompt. We will report it as such.

Beating StateMem means two things on the nominal conditions. First, the form-aware variant must close at least half of the verbal-minus-nominal accuracy gap. Second, it must not lose more than one point on the verbal condition. Both are judged with bootstrap confidence intervals over timelines.

Why not just…?

Why not just give the model the whole history in a long context? The StateMem study shows models with full history fall into the salience trap, and BEAM reports contradiction resolution as the weakest ability even at its scales. Neither result, however, says whether long context fails more on nominal forms. Our long-context and oracle conditions measure that directly. If the long-context model shows no form gap, the extraction-bias hypothesis is wrong for current models, and we will say so.

Why not just prompt the extractor to look for noun-phrase updates? This is the real threat. Instruction-following models may handle "interim director" once told to look for it. It is run as the ablation above, on identical data. The form-aware detector earns its place only if it beats this variant, most plausibly on "former X" and vacancy cases, where a prompt tends to re-extract the stale attribute as a current one.

Why not just stratify errors on existing benchmarks like StateMemBench or BEAM by surface form? It is cheaper, and we will do it as a pilot on a few hundred annotated failures. But natural nominal updates differ from verbal ones in content, entity type and position, so any correlation is confounded. Only matched renderings of the same timeline isolate form. The pilot also serves the opposite purpose: it checks that the effect we find synthetically appears in natural text.

Why not just synthesize the narratives fully automatically and check them with a script? SUMIE's audit shows how often unguided synthesis of evolving attributes goes wrong. Whether "the then-chair" entails a change is a pragmatic judgment, not a string match. We will quantify the need for humans by reporting what fraction of slots that pass automatic entailment checks humans reject. If that fraction is negligible, later versions can drop the human step.

Proposed Experiments

The first decisive check is the form gap itself. We measure current-state accuracy, stale-answer rate and vacancy recognition for every memory system across the five renderings and three length strata. The gap must appear under oracle retrieval before we attribute it to reading rather than retrieval.

The second check tests the salience prediction. If "former X" renderings yield higher stale-answer rates than clean nominalizations at equal distractor counts, lexical re-exposure of the old state is a mechanism, not just nominal parsing difficulty.

The third check is the intervention. The form-aware front end must close the gap relative to StateMem and to the prompt-only ablation. It must then transfer to vacancy questions in Event-level Knowledge Editing, where GPT-4 with in-context editing reached only 29.2% event-level reliability. It should also transfer to SUMIE's UPDATE and MERGE tasks, where the best reported models stay below 80.4% F1. For the transfer runs, we evaluate StateMem with and without the front end on those benchmarks' own splits.

The fourth check concerns generation. Re3-style continuations should show fewer SNaC inconsistency errors per thousand words when the drafter is conditioned on form-aware memory rather than on the StateMem baseline. Human scoring on a segmented sample should agree with the trained detector on the direction of the difference.

The study is informative whichever way it comes out. A large, retrieval-independent form gap that a simple ingestion step closes would reframe part of the knowledge-obsolescence problem as an old extraction problem in new clothes. A null gap would remove a plausible but untested explanation of state drift and point effort back toward temporal reasoning.