A Unified Definition of Hallucination, Or: It's the World Model, Stupid

Abstract: Despite numerous attempts to solve the issue of hallucination since the inception of neural LLMs, it remains a problem in even frontier LLMs today. Why is this the case? We walk through definitions of hallucination used in the literature from a historical perspective up to the current day, and fold them into a single definition of hallucination, wherein different prior definitions focus on different aspects of our definition. At its core, we argue that hallucination is simply inaccurate (internal) world modeling, in a form where it is observable to the user (e.g., stating a fact which contradicts a knowledge base, or producing a summary which contradicts a known source). By varying the reference world model as well as the knowledge conflict policy (e.g., knowledge base vs. in-context), we arrive at the different existing definitions of hallucination present in the literature. We argue that this unified view is useful because it forces evaluations to make clear their assumed "world" or source of truth, clarifies what should and should not be called hallucination (as opposed to planning or reward/incentive-related errors), and provides a common language to compare benchmarks and mitigation techniques. Building on this definition, we outline plans for a family of benchmarks in which hallucinations are defined as mismatches with synthetic but fully specified world models in different environments, and sketch out how these benchmarks can use such settings to stress-test and improve the world modeling components of LLMs.

• The paper’s main contribution is a unified definition of hallucination that frames it as misalignment in an explicit world model.
• It introduces formal components—world model, view function, and conflict policy—to operationalize hallucination detection across different tasks.
• The framework guides the design of benchmarks and evaluation methods, separating world-model errors from other types of model mistakes.

A Unified Account of Hallucination: World Modeling as the Core

Historical Progression of Hallucination Definitions

The paper "A Unified Definition of Hallucination, Or: It's the World Model, Stupid" (2512.21577) presents a comprehensive historical trajectory of the concept of hallucination in neural and LLMs (NLMs/LLMs). Initially, the hallucination phenomenon was defined in narrow terms such as ungrounded or unfactual output in neural machine translation (NMT), where generations deviated significantly from the input source, often through severe omission or fabrication. The definition expanded in abstractive summarization to include spans not supported by source documents, introducing distinctions such as intrinsic vs. extrinsic hallucinations, but still fundamentally referencing the originating document as the primary world model.

As LLMs progressed in scale and ability, focus shifted to broader, world-centric notions of factuality. Hallucination now encompassed any generated content not consistent with an (often) implicit set of world facts, even when explicit sources were absent, as observed in open-domain QA or commonsense generation settings. More recently, the field has pushed toward fine-grained detection, with benchmarks and taxonomy-driven approaches labeling contradictions, entity errors, relation errors, and unverifiable claims, always specified against some reference evidence.

This definitional evolution also manifested in agentic and multimodal contexts. In agentic settings, hallucination addressed incorrect environment modeling that led to invalid actions, and in multimodal systems—especially vision-LLMs—the field moved toward considering parametric-prior and perceptually-grounded conflict as the crux of hallucination (e.g., language priors overtaking visual input).

Formalizing Hallucination as World Model Misalignment

The key contribution of the paper is the explicit unification of hallucination under the lens of a reference world model. This formalization recognizes that hallucination is not a singular error type, but rather any observable output of a model that is inconsistent with a prespecified, context-dependent reference world, as shown schematically below.

Figure 1: Hallucination as inaccurate (internal) world modeling.

The central formal definitions are as follows:

• Reference world model ($W$): A tuple that specifies sets of world states ($\mathcal{S}$), interaction histories ($\mathcal{H}$), and admissible relations ($\mathcal{R}$).
• View function ($V$): Determines which part of $W$ is visible for the task/input.
• Conflict policy ($P$): Dictates resolution when multiple sources provide contradictory information (e.g., should a KB override in-context facts?).
• Truth function ($T_{W,P}$): Labels atomic factual claims as true/false/unknown with respect to $W$ and $P$.

A hallucination occurs if and only if a model output $y$ expresses an atomic claim $c$ such that $T_{W,P}(x, c) = \text{false}$. This criterion operationalizes hallucination detection for any controllable environment and explicitly disentangles it from errors of planning, reward misalignment, or control which do not involve incorrect world beliefs.

Illustrative Cross-Domain Instantiations

The generality of the definition is validated through detailed examples spanning summarization, open-domain QA, retrieval-augmented generation, and agentic browser use. In every case, the instantiation of $W$, $V$, and $P$ yields a precise test for hallucination. For example, in document summarization, hallucination is a function of contradiction to the input document as $W$, with the complete document provided via $V$ and the policy $P$ granting supremacy to the document content. In retrieval-based settings, $P$ may prioritize retrieved context over parametric knowledge, and in agentic (interactive) settings, the observed environment (e.g., a DOM) is the sole ground truth.

This careful decomposition clarifies ambiguous scenarios. For instance, asserting an environment element that does not exist is a hallucination with respect to the observed DOM state, whereas taking a valid-but-suboptimal action is merely a planning error.

Implications for Benchmark and Evaluation Design

A notable implication is the prescriptive guidance on hallucination benchmark design. Any evaluation or mitigation technique must specify (1) the reference world model; (2) the visible subset of that world per input (the model's "observations"); and (3) the conflict policy. This prescription unifies evaluation for otherwise disparate paradigms (summarization, QA, agentic tasks, multimodal reasoning) and enables the principled design of datasets and metrics grounded in observable worlds rather than subjective notions of plausibility.

Furthermore, the authors argue for a paradigm wherein synthetic but fully specified worlds (e.g., games, simulated environments) function as testbeds for large-scale automatic hallucination evaluation. In these settings, the ground truth world is programmatically accessible, and hallucination detection for any model output can be automated through direct comparison with world state and rules.

Figure 2: Overview of the chess scenario environment and instances.

The chess environment is presented as an archetype, where model outputs (e.g., analysis, move suggestions, positional claims) can be decomposed into atomic statements and systematically checked for world-consistency versus the canonical board state and history.

Practical, Conceptual, and Theoretical Benefits

The formalism asserts several strong claims:

• Hallucinations should not be labeled in the absence of a specified $W$, $V$, and $P$.
• The proposed framework is strictly more general and unifying than prior domain-specific definitions.
• Many extant benchmarks and mitigation approaches, implicitly or explicitly, instantiate this template with task-specific choices; making these explicit enables meaningful comparison and meta-analysis of techniques.

Additionally, by separating world model errors from other categories (planning/control/incentive errors), this definition forestalls the risk of "hallucination" becoming a catch-all label for every form of model error. The paper argues that perfect world modeling is unattainable, but models should aim to recognize the boundaries of their knowledge and appropriately abstain when uncertain, which is in line with recent advances in abstention and uncertainty estimation.

Figure 3: Misaligned move—model's prediction vs. ground truth in the chess environment.

Pathways for Future Research

The paper delineates multiple theoretical and practical extensions:

• Varying conflict policies ($P$): Investigating model robustness to explicit knowledge conflicts, adversarial context, or altered retrieval reliability.
• Temporal/world dynamics ($W$): Probing models in non-stationary or evolving environments, relevant for knowledge updating and memory studies.
• Richer synthetic environments: Generalization to dynamic web, code, database, and multimodal simulation environments, leveraging controlled observability and difficulty.
• Distinction between belief and planning errors in agentic settings: Explicit agentic benchmarks where the evaluation surface separately scores world consistency, action validity, and overall task success.

This research agenda aims to both clarify the definition of hallucination across the AI community and provide a scalable framework for rigorous, reproducible, and cross-domain benchmark construction.

Conclusion

By formalizing hallucination as observable output-level divergence from an explicit reference world model, the paper provides a unifying theoretical and methodological framework for ongoing and future research. This approach standardizes terminology, improves benchmark validity, and advances nuanced understanding of model limitations distinct from reward or planning errors. The implications for robust LLM and VLM deployment, especially in dynamic and interactive scenarios, are profound; ongoing expansion to richer task families and evolving environments is both feasible and essential under this paradigm (2512.21577).