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

TL;DR

The paper proposes a unified definition of hallucination as inaccurate internal world modeling observable through outputs, formalized via a reference world model $W$, a view function $V$, a conflict policy $P$, and a truth function $T_{W,P}$. It surveys historical and contemporary notions of hallucination across translation summarization open domain QA RAG and agentic multimodal contexts, arguing that all cases reduce to world model mismatches under different ground truths. The authors provide a practical benchmark blueprint using synthetic worlds and a general three step recipe involving document generation queries and view dependent evaluation, exemplified by a chess domain. They discuss implications for benchmark design mitigation strategies and outline future work including exploring conflict policies changing world knowledge richer environments and separating belief from control in agentic settings.

Abstract

Despite numerous attempts to solve the issue of hallucination since the inception of neural language models, it remains a problem in even frontier large language models 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 language models.

Figures (3)

• Figure 1: Hallucination as inaccurate (internal) world modeling
• Figure 2: Overview of the chess scenario environment and instances
• Figure 3: Misaligned move - the model's prediction (left) vs. reality (right)

Theorems & Definitions (2)

• Definition 1: Reference world model
• Definition 2: Hallucination as inaccurate world modeling