Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

DialFork Benchmarks Procedural Hallucination in Conversational Agents

AI summary

Idiap Research Institute and Logitech released DialFork, an open benchmark for measuring procedural hallucination in conversational LLM agents, including failures where agents remain fluent and use tools correctly but omit or misorder required conversational steps. It includes more than 36,000 dialogues spanning 24 tasks and 12 domains, with drift annotations checked by 30 human annotators and an LLM judge calibrated against them. The data, code, paper, predictions, human judgments, and three task leaderboards are publicly available.

In-depth insight

DialFork adds procedural adherence as a distinct evaluation axis alongside factual accuracy and correct tool use, giving agent developers a public test set and leaderboards for measuring a failure mode described in the release.

Hi all, author here (Idiap Research Institute, joint work with Logitech). We recently released DialFork and thought it might interest people here. It's a benchmark for a failure mode of LLM agents that is easy to miss: every turn is fluent and factual, the tool calls are correct, yet the agent stops following the procedure it was given. For example, asked for the weather, it fetches the forecast (correct tool call) but never tells you, and starts asking you to rate its service instead.

Most steps of a conversational procedure aren't tool calls but things the agent has to say: asking, confirming, telling the result. When an agent drifts from these steps, we call it procedural hallucination: a third axis, alongside factual accuracy and correct tool use, and one that has been largely overlooked. To our knowledge, DialFork is the first benchmark aimed at measuring this axis: what the agent has to say, and in what order.

How it's built: we simulate dialogues where an agent follows its workflow (24 tasks, 12 domains), then fork them: we replay each one and, at a chosen turn, give the agent an instruction that makes it drift, in different, complementary ways. That gives over 36k dialogues where we know where each drift starts and how far it goes, checked with 30 human annotators and an LLM judge calibrated on them.

#DialFork#Idiap Research Institute#Logitech#Hugging Face

Original · 3 min readRead the original
Group chat

No comments yet. Start the conversation.