DialFork Tests Whether Agents Follow the Conversation Procedure

Introduction

Ask an agent to check the weather and it may remain polite, coherent, and free of obvious factual errors—while never actually delivering the forecast. The failure is not necessarily a false statement. It is the agent’s inability to keep the conversation moving through the procedure required by the task. The DialFork benchmark targets this overlooked failure mode, which the paper calls procedural hallucination.

What DialFork measures

• Process rather than isolated turns: Conventional factuality and response-quality metrics often score each answer separately. They may miss skipped steps, inappropriate transitions, or conversations that remain plausible but make no progress toward completion.
• Controlled dialogue forking: DialFork starts with dialogues that follow the intended procedure. At controlled turn positions, the researchers introduce a change so that the continuation forks away from the conformant path. The fork location supplies a useful signal for identifying when deviation begins, reducing reliance on exhaustive manual annotation.
• Broad benchmark coverage: The dataset contains 36,278 English dialogues across 24 tasks and 12 domains, with four types of procedural hallucination. Thirty human annotators were used to validate the construction.
• Three evaluation settings: Set-to-set distance supports monitoring of deployed agents. Dialogue-to-dialogue distance supports procedural retrieval, such as finding conversations that follow a similar path. Turn prediction aims to identify the first off-track turn for real-time guardrails.
• Flow-aware metrics: FlowKL and FlowScore compare how dialogues unfold instead of treating semantic similarity as a proxy for procedural similarity. According to the paper’s summary, FlowScore performs better than the evaluated baselines at the dialogue level across seven encoders, reinforcing the distinction between similar meaning and similar process.

Why it matters

The distinction is important for customer support, booking, information collection, and tool-using agents. A response can be locally acceptable while the overall interaction is operationally useless. DialFork gives researchers and developers a way to measure whether an agent is following the expected route, and whether a guardrail can intervene near the first point of failure rather than waiting until the conversation has fully derailed.

The benchmark should not be interpreted as a complete definition of dialogue success. Users can change goals, and a capable agent may need to adapt its plan when new information appears. A deviation from a reference path is therefore best treated as a monitoring or diagnostic signal, alongside task completion, user satisfaction, and human review.

By releasing DialFork’s dataset and code under the MIT license, the authors provide a starting point for evaluating both what an agent says and how it conducts a task over time. That shift could help move conversational-agent testing beyond sentence-level correctness toward process reliability.

Source: Hugging Face Daily Papers