anonymous-neurips2026-edsubmission/

SGR-Bench

SGR-Bench is a dataset for evaluating State-Gated Retrieval. In this setting, the answer-bearing evidence is not exposed in the default state of a website. It becomes available only after the user configures site-internal filters, views, hierarchies, scopes, time windows, or result pages. This release contains 100 aligned task records organized into two prompt formulations:

goal: a goal-oriented version that states the target information need while leaving more of the… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-neurips2026-edsubmission/SGR-Bench.