DialFork Benchmark for Procedural Hallucination Detection
October 14, 2026
DialFork is a new benchmark of 36,278 dialogues designed to measure procedural hallucination, where agents follow fluent but logically incorrect conversational processes. It identifies failures in following intended conversational steps rather than just factual errors.
HOW THIS AFFECTS YOU
●
builderYou should test your conversational agents for logical flow, not just factual accuracy.
●
researcherThis provides a way to quantify how agents deviate from complex, multi-turn instructions.