Pooled recall is the total number of unique vulnerabilities found across multiple runs of the same model. It captures the benefit of repetition when different passes surface different issues. This metric is useful when assessing search-oriented security workflows where one run alone does not represent full capability.
What Pooled Recall Measures
Pooled recall is a coverage metric for repeated model runs. It counts the total number of unique vulnerabilities surfaced across multiple passes, so the emphasis is not on one individual run but on the cumulative discovery achieved when a workflow is repeated.
That makes it useful for judging search-oriented or exploration-heavy security workflows, where a single execution may miss issues that appear only on later passes.
Why Repetition Changes the Meaning of the Metric
The key idea behind pooled recall is that repeated sampling can uncover different findings because the model’s search path is not perfectly stable. A model that appears modest on one run may still have strong aggregate discovery capability if later runs reveal additional distinct vulnerabilities.
This is why pooled recall is best read alongside per-run performance. A high pooled recall with weak single-run recall suggests the workflow benefits from repetition, reranking, or multiple prompting strategies rather than from a single deterministic pass.
How to Interpret It in Security Evaluation
For vulnerability discovery tasks, pooled recall helps answer a practical question: how much of the issue space can this workflow find if I allow multiple attempts? That makes it especially relevant when the goal is breadth of discovery, not just consistency or precision.
It also helps compare search systems that may trade stability for exploration. Two workflows can look similar on one run, yet differ materially once you measure the unique vulnerabilities they surface across several runs.
For teams evaluating AI-assisted security workflows, the metric is most meaningful when the underlying test set, run count, and deduplication rule are held constant. Without that discipline, pooled recall can look larger simply because the workflow was allowed more opportunities to search.
Limits of the Metric
Pooled recall does not tell you whether the model found the same issues reliably, whether the findings were ranked well, or whether the workflow is operationally efficient. It also does not substitute for precision, because a system can accumulate many unique findings while still producing a noisy output set.
Used well, it is a measure of aggregate discovery potential. Used badly, it can overstate practical capability by hiding how much repetition was required to reach that total.
Practitioner Guidance
Why practitioners should care: Pooled recall is useful when the security task rewards exploration, but it should be interpreted as a cumulative discovery metric rather than a single-pass quality score. If you use it in evaluation, make sure stakeholders understand that it favors workflows that keep finding new issues over repeated attempts.
Common misunderstanding: A higher pooled recall does not automatically mean the model is better in practice. It may simply mean the system needs multiple runs to expose the same underlying capability, so you still need complementary measures for reliability and usefulness.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 19, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org