Anonymous research manuscript for evaluating the pragmatic decision that precedes answer generation in retrieval-augmented generation.
Beyond Answer or Abstain: Human-Validated RAG Action Selection
- Manuscript: askrag-bench-anonymous.pdf
- Status: Anonymous research submission
Most retrieval-augmented generation evaluations begin after a system has already chosen to answer. AskRAG-Bench evaluates the earlier, pragmatic choice among four mutually exclusive actions: ANSWER, CLARIFY, ABSTAIN, or QUALIFY.
The resource separates a source-grouped construction pool from target-blind human validation and a hash-reconstructable, high-consensus evaluation core. It also provides controlled same-record contrasts that preserve source identity while changing an action-relevant question or evidence condition. The benchmark supports reproducible analysis of when a RAG system should answer, ask for clarification, abstain, or provide a qualified response.
askrag-bench-anonymous.pdf: Anonymous manuscript.README.md: Publication overview.
This repository is maintained as an anonymous research release. It intentionally contains no identifying author, affiliation, or contact information.
Citation metadata will be provided after the anonymous review process is complete.