Observed arrival · 2026-09-26
HalluWorld maps where language models still make things up
HalluWorld is a benchmark for measuring language-model hallucinations across grid worlds, chess, and realistic terminal tasks.
- For
- Researchers evaluating language-model reliability
- Worth noticing
- Its probes distinguish perceptual, causal, memory, and uncertainty errors across grid worlds, chess, and terminal tasks.
Field notes
The framework defines a hallucination as an observable claim that is false in a fully specified reference world. Its homepage says the environments allow controlled variation in world complexity, observability, temporal change, and source-conflict policy. Reported results distinguish near-solved perception from harder state tracking, forward simulation, and abstention; the page also cautions that extended thinking does not generally solve the causal probes.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue