Grounding-Blind Judges: Multi-Judge Evaluation of Scientific RAG Reveals Systematic Divergence between Subjective Scores and Citation Validity
Retrieval-Augmented Generation (RAG) systems in scientific domains promise verifiable citations, yet dominant evaluation frameworks rely on LLM judges without external grounding. We present REVAL (Retrieval-Verified Evaluation), a multi-judge evaluation framework anchored in CrossRef citation verification, and BodybuildingQA, an exercise-science question-answering benchmark (350 questions, Brazilian Portuguese). REVAL pairs cross-judge scoring with Verified Citation Accuracy (VCA), which separates whether a cited DOI exists from whether it matches the cited author and year. VCA verifies citation metadata, not claim support. Evaluating a closed-source RAG system against four generalist LLMs and two web-search conditions, we find that, in this domain and language, judge scores and metadata-verified citation accuracy diverge: judges rank the system with the most verifiable citations last, while rewarding fluent, DOI-rich answers whose citations are frequently fabricated or misattributed. In a controlled perturbation study (identical answers, only DOI strings replaced by non-resolving fakes), judges are severely under-sensitive to citation validity: format-preserving fakes lower their Citation Verifiability score by only 0.21 points (1–5 scale), a ninth of their response to format-breaking fakes (1.91). Web search, tested in two provider families, halves fabrication yet leaves a substantial metadata-verification gap and does not close the divergence. We release the benchmark, judge scores, and scripts. Accepted to EACL 2027 (Main Conference). This is the camera-ready version.
Authors
- Bruno P. Roquette (ORCID: https://orcid.org/0000-0002-0071-2361)
- Alexandre Cristovão Maiorano
Institutions
- Lumina (Sweden) (SE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24407411
- Primary Topic
- Topic Modeling
- Type
- preprint