If your AI agent can fetch the weather but then forgets to tell you the forecast, instead asking for a satisfaction rating, you have encountered something more subtle than a factual error or a broken tool call. This is procedural hallucination, and DialFork, a new benchmark from Idiap Research Institute and Logitech, exposes it as a critical blind spot. The agent says the right words and uses the right tools, yet abandons the actual procedure it was given. For anyone building or relying on AI agents, this is not an edge case, it is a fundamental missing dimension of reliability. As The Open Source AI Boom Demands a New Answer to an Old Question: Who Wins explores, the ecosystem is racing toward more capable models, but capability without procedural fidelity is a trap. Meanwhile, A decade of ML research distilled into one clear page per topic reminds us that clarity in evaluation is the foundation of progress, and DialFork brings that same clarity to a previously murky problem.
Most benchmarks measure what an agent knows or whether it can call a function. DialFork measures whether the agent follows the script of an interaction: asking before acting, confirming before proceeding, reporting results before moving on. The researchers generated over 36,000 dialogues, then deliberately inserted instructions that cause the agent to drift from its procedure. The result is a dataset where every drift is labeled and measured, from the exact turn it starts to how far it travels across the dialogue. This is not abstract. If you are monitoring agents in production, Task 1 lets you compare entire sets of conversations to see if a model upgrade makes agents more prone to procedural drift. Task 3 offers a live guardrail, detecting the moment an agent goes off-script. These are tools for engineers who need to trust that their agents will complete a workflow, not just answer a query.
The practical takeaway is direct: your current evaluation pipeline is likely blind to procedural hallucination. You can test factual accuracy and tool call correctness, but if your agent completes a booking without confirming the date, or fetches a report without reading it aloud, you will not catch it with standard metrics. DialFork opens a new axis for evaluation, one that treats the order and completeness of an agent's utterances as a measurable property. Explore simpler ways to open CSVs without endless extra steps speaks to the value of removing friction from data tasks; DialFork does the same for agent evaluation, offering a clean benchmark and a leaderboard on Hugging Face that accepts predictions as a single CSV file. The barrier to entry is low, and the room for improvement is wide.
The open question is whether procedural hallucination is a model limitation or a training data artifact. DialFork gives the community a way to find out. The leaderboard tracks three tasks at increasing levels of granularity, and the initial results leave plenty of headroom. For anyone building agents that handle real tasks, customer support, data entry, scheduling, this is the benchmark to watch. Not because it is the flashiest, but because it measures something your users will notice the moment it goes wrong.