Benchmarks / PQR

PQR

Released
QA-agent failure discovery

Generating diverse and realistic user queries that are specifically useful for exposing QA-agent failures.

Updated May 2026

Background

Average-case question sets can conceal important reliability problems. PQR focuses query generation on realistic questions that reveal where a deployed QA agent breaks.

What it evaluates

  • Diversity and realism of generated user queries
  • Ability to uncover previously missed agent failures
  • Coverage of difficult behaviors and edge cases
  • Usefulness for iterative QA-agent improvement

Evaluation methodology

Generated queries are run against a target QA agent and assessed by the failures they elicit, not simply by surface diversity.

Representative task

Generate a realistic query for a target QA system that probes a meaningful capability boundary and reveals an actionable failure.

Results

The paper reports the generation framework, evaluation protocol, and comparative results.