110 scenarios testing whether a model can write orchestration prompts that give each sub-agent in a multi-agent system exactly what it needs to know, without leaking irrelevant context.
unassessed
| Category | agentic |
|---|---|
| Subcategory | multi-agent orchestration prompt composition (role-fragment assignment and free-form prompt writing) |
| Page status | active |
| Metric | Strict-pass rate (plus a separate information-leakage count) |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 220 |
| Dataset licence | MIT, per the Hugging Face dataset card |
| Publisher | University of Maryland, with co-authors at the Chinese University of Hong Kong and Stanford University |
Despite a name generic enough to suggest something about human perspective-taking or opinion polarisation, PerspectiveGap measures a narrower and more specific capability: whether a model can compose orchestration prompts for a multi-agent system by correctly working out what each sub-agent role actually needs to know, and, just as importantly, what it should not be told. Given a scenario describing a multi-agent workflow, the model must either assign the correct information fragments to each role (with distractor fragments mixed in) or write the orchestration prompt text itself from scratch. The 110 scenarios are organised into 10 orchestration "topologies" the authors say are distilled from their own real-world multi-agent engineering practice, and the benchmark is framed around what the paper calls the "Prompt Economy" principle: building orchestrations that reuse a small number of stable roles rather than proliferating roles and handoffs.
Two distractor-mixed task formats scored separately and then averaged: role-fragment assignment (selecting which information fragments belong to which sub-agent role, out of a set that includes at least one injected distractor) and free-form prompt writing (generating the actual orchestration prompt text for a role). A separate ablation varies the injected-distractor count from 0 to 3.
No model card in ModelSpec reports this benchmark yet.