HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models
Emmy Liu, Varun Gangal, Michael Yu, Zhuofu Tao, Karan Singh, Sachin Kumar, Steven Y. Feng
cs.CL, cs.AI, cs.LG, stat.ML
2026-05-19
HalluWorld benchmarks hallucination in controllable worlds; frontier models nearly solve perception, but causal errors persist and extended thinking makes them worse.
Hallucination in large language models is an old problem, but evaluation is fragmented. Summarization, QA, RAG, and agentic tasks each have their own hallucination tests, built on incompatible definitions, so a mitigation that works in one setting tells you nothing about the next. Existing benchmarks either lean on human annotation (which models will eventually memorize) or observe naturalistic settings (which cannot be reproduced systematically).
HalluWorld's goal is to define hallucination inside a controllable, automatically scorable framework, so you can twist variables like world complexity, what the model observes, and which source to trust under conflict, and measure each failure mode separately.
The backbone is a reference-world formulation: a model hallucinates when it emits a claim that is false with respect to a reference world W. Four components define it, the world W (states and transitions), a view function V (what the model sees), a conflict policy P (how contradictory information is resolved), and a truth function T (correctness).
All three environments are programmatically generated and auto-scored: gridworlds (MiniGrid, 33 levels, 839 probes), chess (Lichess positions, 7 levels, 350 probes), and terminal tasks (real software-engineering tasks from Terminal-Bench, 110 tasks, 529 probes). Probes fall into five categories targeting distinct cognitive demands: Perceptual, Memory, Causal, Uncertainty, and Compound. A dozen-plus frontier and open models are benchmarked.
On perceptual probes, frontier models are near-perfect. On gridworlds GPT-5.5 has an overall hallucination rate of 5.0% (perceptual 0.4%), Claude Opus 4.6 is at 5.9%, and o3 at 6.8%. This category is essentially solved.
Causal reasoning and uncertainty are still bad. Kimi K2.6 sits at 9.3% overall but 47.4% on causal probes. On terminal tasks, uncertainty is the hardest probe (GPT-5.5 still at 23.1%).
The most counterintuitive finding is that turning on extended thinking makes causal hallucination worse. Sonnet 4.6's causal rate climbs from 19.7% to 24.0%, and Opus 4.6's from 20.1% to 25.2%. Thinking longer, on tasks that require simulating the world forward, makes the model more likely to err.
| Model | Grid overall | Perceptual | Causal |
| GPT-5.5 (T) | 5.0% | 0.4% | 25.3% |
| Claude Opus 4.6 | 5.9% | 3.7% | 20.1% |
| o3 | 6.8% | 0.3% | 26.0% |
| Kimi K2.6 | 9.3% | 5.4% | 47.4% |
The value here is not a leaderboard. It is a decomposition of hallucination into independently measurable failure modes. For model builders, "perceptual is nearly solved, causal and uncertainty are far from it" is a clear research map. The thinking-makes-it-worse result is the one to flag: better reasoning and less hallucination are not the same thing, and a longer chain of thought can be a liability in settings that demand forward simulation.
Probes measure only the beliefs a model states explicitly, not its internal representations. Probe locations are co-designed with the environments rather than discovered via active learning at points of maximal cognitive load. The terminal setting is complex enough that isolating individual failure modes is harder than in synthetic ones. Model rankings swing wildly across probes (on one level Opus and Sonnet score 0% while o3 hits 43.3%), so the overall ranking is highly sensitive to probe mix; do not read much into a single aggregate.