Source: arXiv — 2026-07-09

Summary

A new paper benchmarks whether smaller, cheaper rubric-based LLM judges can replace frontier models specifically for citation and source-attribution verification in deep-research pipelines — the step that checks whether a generated claim is actually backed by the source it cites — and identifies specific bias patterns in how these judge models evaluate attribution.

Key Takeaways

• Targets a narrow, high-value question: for the specific sub-task of checking whether a citation actually supports a claim, does a full frontier model add value over a smaller rubric-based judge?
• Identifies specific judge biases in rubric-based citation verification, meaning cheaper judges don't fail randomly — they fail in predictable, characterizable ways.
• Directly relevant to anyone running a deep-research or RAG pipeline at scale, where citation-verification cost (running a frontier model as a judge on every claim) compounds fast.
• Extends the digest's recurring LLM-as-judge coverage (bias-vs-noise tradeoff research already logged) into the specific, practical citation-verification use case rather than judge evaluation in the abstract.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)