A Controlled Benchmark for Hallucination via Reference World Models
A model hallucinates when it makes an observable claim that is false in a fully specified reference world.
Frontier models rarely hallucinate about directly observed state: seven of twelve models reach 0.0% perceptual error in grid worlds, and five reach 0.0% on chess with no FEN provided.
Multi-step state tracking and forward simulation remain hard: every model hallucinates on at least 16.7% of grid memory probes, and even the best models fail 18–32% of causal chess probes once an incorrect FEN is in the context.
Models struggle to say "I don't know": no model gets below 24% hallucination on grid uncertainty probes, and in realistic terminal tasks uncertainty stays at 9–27% while other categories drop near zero.
Hallucination remains a central failure mode of large language models, but existing benchmarks operationalize it inconsistently across tasks such as summarization, question answering, retrieval-augmented generation, and agentic interaction. This fragmentation makes it unclear whether a mitigation that works in one setting actually reduces hallucinations across contexts. Current hallucination benchmarks either require human annotation and fixed references that may eventually be memorized, or rely on naturalistic observations often recorded in settings that are difficult to reproduce or test systematically.
To enable further research on the root causes of hallucination, we introduce HalluWorld, an extensible benchmark framework grounded in an explicit reference-world formulation: a model hallucinates when it produces an observable claim that is false with respect to this reference world. Building on this view, we construct a family of synthetic and semi-synthetic benchmark environments in which the reference world is fully specified, the model's observable view is controlled, and hallucination labels can be generated automatically by construction. HalluWorld spans multiple settings that are classically representative for AI — grid worlds, chess, and realistic terminal tasks. This enables controlled variation of key factors such as world complexity, observability, temporal change, and source-conflict policy, allowing us to disentangle hallucinations into more fine-grained error categories.
We evaluate frontier and open-weight language models across these settings and find consistent patterns across domains: perceptual hallucination on directly observed information is near-solved for frontier models, while multi-step state tracking and causal forward simulation are still difficult, and are not generally solved by extended thinking. In the terminal setting specifically, models also struggle with when to abstain from answering. The uneven profile of failures across probe types and domains suggests that different hallucinations arise from qualitatively distinct failure modes rather than reflecting a single underlying capability. Our results suggest that controlled reference worlds offer a scalable and reproducible path toward measuring and reducing hallucinations in modern language models.