Deep-Research Citation Benchmark - Source Relevance: leaderboard
The Deep-Research Citation Benchmark (PwC, 2026) uses LLMs as rubric judges of whether a cited web source is topically relevant to the claim it supports, over 624 attribution-citation pairs from an adversarial long-form research report. The score is Cohen kappa against human-reviewed gold labels, from -1 to 1; higher is better.
Metric: Cohen kappa (-1 to 1) between the LLM judge and human-reviewed gold labels on the source relevance rubric criterion of 624 attribution-citation pairs parsed from an adversarial 25-domain deep-research report (scores binarized at 0.5); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 0.64 | #348 |
| 2 | Claude Opus 4.6 | 0.55 | #102 |
| 3 | GPT-5.4 Mini | 0.54 | #389 |
| 4 | GPT-OSS-120B | 0.52 | #700 |
| 5 | Gemini 3.1 Pro (Preview) | 0.46 | #94 |
| 6 | Gemini 3.1 Flash Lite (Preview) | 0.46 | #418 |
| 7 | Claude Haiku 4.5 | 0.34 | #568 |
| 8 | Claude Sonnet 4.6 | 0.32 | #161 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=deep-research-citation-benchmark-source-relevance · How It Works · Data refreshed daily, snapshot 2026-11-13.