DialFork Benchmark Evaluates Procedural Hallucination in Conversational Agents
October 16, 2026
DialFork introduces a benchmark for measuring procedural hallucinations, where agents follow fluent logic but deviate from intended conversational processes. The dataset contains 36,278 dialogues across 24 tasks and 12 domains to evaluate deviation granularity.
HOW THIS AFFECTS YOU
●
builderYou can use this to test whether your conversational agents maintain logical process flow beyond simple factuality.
●
researcherThis provides a standardized way to study non-factual model failures in multi-turn dialogues.