synthetic-environments

HalluWorld is a controlled benchmark framework for evaluating hallucination in large language models using explicit reference world models across synthetic environments like gridworlds, chess, and realistic terminal tasks. It enables fine-grained analysis of failure modes such as perceptual hallucination, multi-step state tracking, and causal simulation, revealing that frontier models still struggle with complex reasoning not solved by extended thinking.