Planner-Labeled Hallucinations in Goal-Driven Stories

When a language model writes a long story, it makes mistakes that no fact-checker can catch. A character opens a vault without the key they never obtained. A character reacts to a secret they could not have witnessed. A character abandons the goal that drove the last ten pages for no reason. Hallucination research has built detailed error categories for claims about the real world and for tool-using agents, but it has no ground truth for fiction, where nearly every statement is "invented" by design. As a result, nobody can say where a story first broke, or which kind of break it was. Nobody can say whether a detector that flags a character's odd choice is catching a bug or punishing the drama.

We propose to build that ground truth using an old tool from AI storytelling: a narrative planner. This is a program that assembles a story from actions, each with explicit preconditions, a motivating character goal, and a record of which characters witnessed it. A language model turns each planned step into prose. It then continues the story freely, and we map its continuation back onto the planner's state. The planner's own bookkeeping labels every violation automatically, with a type and the step where it began. It also marks deliberate authorial deviations, such as a character acting against self-interest for suspense, as licensed. This gives the first benchmark for attributing hallucinations in stories that separates broken intentional causality from dramatically useful irrationality. It also lets us test whether typed, scene-level feedback repairs stories better than the single-score control used in long-story generators today.

Core Claim

The claim is that a planner-grounded story world makes narrative hallucination both attributable and type-separable in a way that human annotation of free-form stories cannot match at scale. A benchmark built this way should expose one specific weakness in current language models. They fail to localize where an intentional-causality error started, and they cannot tell such errors apart from licensed deviations. We expect this weakness to persist when we measure it on naturally generated long stories rather than synthetic ones. If it holds, the practical consequence is that repair feedback should name the error type and origin scene. A scalar "coherence" score is not enough.

The file contains no direct precedent for this combination. Explore Theory of Mind is the closest piece of machinery. It searches over a hand-built action language that tracks world state and each character's beliefs, then renders prose with a language model. Its characters, however, have no goals, and its evaluation is multiple-choice questions about beliefs, not errors made during generation. We keep its state-tracking and prose-rendering approach, but we add intentional structure and evaluate the generation itself. AgentHallu defines the task we transfer, which is finding the step responsible for a hallucination and explaining it. Its labels are human annotations of tool and web agents judged against external ground truth. Our labels come from the planner and are judged against the story's own internal world. Fine-grained Hallucination Detection and Editing for Language Models (FAVA) supplies the precedent for training on typed synthetic errors. It explicitly excludes commonsense and causal errors, and those are exactly the errors we target.

Research Program

The story engine. We extend the action language of ExploreToM with the intentional machinery that narrative planning formalized, as reviewed in The Story So Far on Narrative Planning. In that framework every character action must belong to a sub-plan serving that character's goal, and an unmotivated action counts as an unresolved flaw. Each generated step therefore records four things: its physical preconditions, the goal it serves, the characters who know it happened, and an expected emotional reaction. The emotional reaction draws on the motivation-to-action-to-emotion links that COMMA formalized using Maslow's needs and Plutchik's emotions.

The planner can also emit licensed deviations. These are steps that violate one character's rationality but serve an explicit authorial goal, such as suspense, withheld information, or a deception whose target lacks knowledge. The planner annotates each one with its justification. The file notes that expanding such action languages requires manual authoring. We accept that cost for roughly a dozen settings, and we treat broader coverage as a secondary goal rather than a claim.

Natural errors, planner labels. Injecting corruptions would teach detectors to spot artifacts of the injection. Instead, we render a plan prefix into prose and then ask a language model to continue the story for several scenes toward the stated character goals. An extraction model maps each continued sentence to planner actions. The planner then checks each extracted action against the current state. An action with an unmet precondition is a prerequisite hallucination. An action that serves no active goal of the character, or that contradicts that goal, is a motivational inconsistency. An action that depends on information the character never received is an epistemic hallucination. An emotional reaction that clashes with the need that was just satisfied or frustrated is an affective disconnect.

The origin step is the earliest sentence whose extracted action introduces the inconsistent state. This lets us separate root errors from later errors that follow from them. We record errors in all four categories, but we treat affective disconnect as the least reliable category, because its reference model is a psychological scheme rather than logic.

The weak link is extraction from prose. Hayton et al. show that recovering planning models from text is hard. We will therefore report extractor accuracy against human annotation on a stratified sample. Any category the extractor gets wrong often enough to distort the benchmark will be dropped.

The typed scene-level critic. On the resulting data we train a mid-sized critic model. It reads a story up to the current scene and outputs, for each scene, an error type, a pointer to the origin scene, and a short explanation, or "licensed" or "clean." This follows the typed-feedback design of SWE-PRM. That work showed that descriptive guidance organized by error category, applied only at intervals, beat prescriptive action suggestions for software agents.

Baselines

The naive reference point is a story generator with no checking, which gives the raw rate of hallucinations by type. For attribution, the strongest existing method is a frontier model given the whole story and prompted with our taxonomy, using AgentHallu's step-by-step incremental protocol. That protocol improved attribution by an average of 12.3% across models, and with it Gemini-2.5-Pro reached 41.1% step localization accuracy on agent trajectories (AgentHallu). This is the strongest competitor because it needs no training and represents the view that "a good judge is enough." We will run it on identical stories with the same few-shot examples and the same budget for tuning the taxonomy description as our critic receives. There is no reported result on Lacuna for narrative hallucination attribution. To beat it, our critic must exceed this judge on origin-scene localization at matched false-positive rates on licensed deviations, on human-annotated natural stories and not only on planner-labeled ones.

The second baseline is an untyped coherence detector trained on the same stories. It uses the sentence-level architecture of Narrative Incoherence Detection, which reached 80.4% on the TripAdvisor discordant-sentence task. It separates the value of the planner data from the value of the error types.

For repair we compare against DOC-style outline control (DOC), which reported a 22.5% absolute plot-coherence gain over Re³ on WritingPrompts. We also compare against CCI, which human raters preferred over DOC 79.5% of the time on ReedsyPrompts (A Character-Centric Creative Story Generation via Imagination). Both use the same drafting model and the same number of regeneration calls.

The key ablation removes the type and origin pointer from the critic's feedback and leaves only a flag or a scalar score. A second ablation removes licensed-deviation examples from training.

Why not just…?

Why not just prompt a frontier model with the taxonomy and skip training? This is the real threat. On short, fully specified stories it may well match a trained critic. The evidence against it is indirect. Localization falls from 29.9% to 11.4% as agent trajectories lengthen (AgentHallu). GPT-4o correctly identifies how many levels of belief reasoning a question requires, yet answers only about 59% of those questions correctly (Mind Your Theory). It scores 9% on adversarial belief-tracking stories (Explore Theory of Mind). If the prompted judge matches the critic across story lengths on natural human-annotated stories, we will report that and drop the critic. The benchmark and the licensed-deviation measurements would then be the contribution.

Why not just prevent errors with better outline control? Prevention and diagnosis are different jobs. DOC's authors still report errors in names, genders, and world prerequisites, and DOC's scalar controller cannot report which constraint drifted. The repair experiment compares both approaches directly on the same drafting model.

Why not a binary coherence or self-consistency check? Self-consistency methods such as Agrawal et al. detect answers that change under cross-examination. Deliberate withholding and deception in fiction are designed to look inconsistent, so such methods should flag licensed deviations. The untyped-detector baseline and the licensed-deviation false-positive rate test this directly.

Why not just retrieve from the story so far? Retrieval grounds surface facts. In FAVA's evaluation, retrieval-augmented ChatGPT reached only 27.8% F1 against FAVA's 48.1%. Beyond that, motivation and knowledge boundaries are not stored in the text as retrievable facts; they are consequences of the story's structure.

Proposed Experiments

The first experiment measures attribution. We report localization accuracy and type accuracy on planner-labeled continuations, broken down by category and by story length measured in scenes. We then compare the length curve with AgentHallu's to test whether narrative attribution degrades in the same way.

The second experiment is the tellability test. It measures each detector's false-positive rate on licensed deviations versus true violations of the same surface form, for example the same irrational act with and without an authorial justification.

The third experiment tests transfer and is the decisive one. We generate long stories on WritingPrompts and ReedsyPrompts with a DOC-style pipeline. Annotators then label origin scenes and types, and we also run them on a 200-story subset of planner data to check that their labels agree with the planner's. The critic must keep its advantage over the prompted judge on these natural stories. Otherwise the planner data taught artifacts rather than narrative causality.

The fourth experiment tests repair. In a scene-level revise loop, typed feedback, scalar feedback, and no feedback are compared on human-rated motivational coherence and on interestingness. Rating interestingness guards against the failure mode in which the repair loop produces stories that are coherent but dull.

The idea succeeds if typed, origin-pointing critique improves coherence without lowering interestingness, and if licensed deviations survive the repair loop at a higher rate than under scalar control.