DialFork: first benchmark targeting procedural hallucination in LLM agents, from Idiap

sergbur · reddit · 2026-11-05

Idiap Research Institute (with Logitech) released DialFork, a benchmark for procedural hallucination — an overlooked agent failure mode where every turn is fluent, factual, and tool calls are correct, yet the agent drifts off the given procedure (e.g., fetching a forecast but never reporting it, then asking you to rate the service).

How it's built: dialogues where agents follow workflows (24 tasks, 12 domains) are replayed and forked at a chosen turn with drift-inducing instructions, yielding 36k+ dialogues with known drift start and extent, verified by 30 human annotators and a calibrated LLM judge.

Three tasks:

Turn level (Task 3): locate the drift turn, streaming-friendly — usable as a live guardrail.
Dialogue level (Task 2): measure procedural distance between a correct dialogue and its hallucinated variants — a new dimension beyond semantic similarity, for retrieval/clustering by how conversations unfold.
Set level (Task 1): a single distributional distance between sets — useful for monitoring agents, e.g., drift after model upgrades.

Baseline results (embeddings, streaming detectors, zero-shot LLMs) leave ample room. Everything is open: data, code, all paper predictions, human judgments, plus a public HF leaderboard where submitting takes one CSV.

Original post →

DialFork: first benchmark targeting procedural hallucination in LLM agents, from Idiap

More from coding & agent