evaluate-task

You can evaluate WebBench task execution traces by applying LLM-based scoring to the execution traces, final page state, and extracted data against the original task description to determine a verdict.

Post-hoc LLM evaluation is a non-interactive analysis process that reviews completed task executions, outcomes, and traces against a rubric to produce structured verdicts suitable for automated QA pipelines.

To get structured verdict reasoning for task verification, you apply LLM-based scoring to the task's traces and outcomes, which generates a PASS, PARTIAL, or FAIL verdict along with the supporting reasoning.

Yes, you can use non-browser evaluation workflows to analyze WebBench tasks by applying post-hoc LLM-based scoring to execution traces and final states rather than running interactive browser sessions.

Yes, LLM-based task evaluation supports partial pass outcomes by analyzing extracted data and execution traces to assign a PARTIAL verdict when a task is only partially completed according to the rubric.

To evaluate a completed WebBench run, you need to provide the original task description, execution traces, final page state, and extracted data so the LLM can apply scoring and generate a verdict.