DialFork: Going Beyond Factuality to Evaluate Procedural Hallucination in Conversational Agents

Key points

• Large language models (LLMs) deployed as conversational agents are increasingly expected to follow procedural logic.1
• While factual hallucination has received extensive attention, a different failure mode has been overlooked: procedural hallucination, in which an agent produces individually fluent and factually plausible utterances while deviating from the intended conversational process.1
• We define three evaluation tasks that characterize procedural deviation at three granularities: how far a set of dialogues deviates from conformant ones (monitoring deployed agents), how far a single dialogue does (procedural retrieval), and at which turn a dialogue first goes off-track (real-time guardrails).1
• For the first two, we propose dialog-flow-aware metrics, FlowKL and FlowScore, that compare dialogues by how they unfold rather than their content.1

Sources (1)

• [1]DialFork: Going Beyond Factuality to Evaluate Procedural Hallucination in Conversational Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 09:18 AM

  We introduce DialFork, the first benchmark to measure this failure mode on raw dialogue text.

  Large language models (LLMs) deployed as conversational agents are increasingly expected to follow procedural logic.

Extractive summary: sentences quoted from the sources.