HiEviDR-Bench - Wiki (RAG): leaderboard
HiEviDR-Bench (UCAS, Tsinghua and others, 2026) tests multimodal deep-research reports on 1,000 Wikipedia questions under the one-pass retrieval-augmented pipeline. A Qwen3-VL-235B-A22B judge scores each report against a hierarchical evidence graph on five 0-20 dimensions (report quality, evidence traceability, and gated citation, claim and answer scores); the score is their sum from 0 to 100, higher is better.
Metric: Overall score (0-100): sum of five 0-20 dimension scores (multimodal report quality, evidence traceability, and progressively gated citation, claim and answer scores) judged by Qwen3-VL-235B-A22B-Instruct against the hierarchical evidence graph, on the 1,000-question Wikipedia subset; MLLM with RAG: one-pass report generation over the top-15 retrieved multimodal evidence items; higher is better. Source: arxiv.org. Saturation forecast: Around 2033. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 42.67 | #30 |
| 2 | GPT-5 | 41.89 | #167 |
| 3 | Gemma 4 31B (IT) | 41.71 | #342 |
| 4 | Grok 4.5 | 40.82 | #82 |
| 5 | GPT-5 Mini | 40.18 | #348 |
| 6 | Qwen 3.5 27B | 39.35 | #359 |
| 7 | gemma-4-E4B-it | 39.16 | #1469 |
| 8 | Qwen 3 VL 8B Instruct | 38.9 | #981 |
| 9 | Qwen 3.5 35B A3B | 37.72 | #482 |
| 10 | Qwen 3.5 9B | 36.07 | #803 |
| 11 | Qwen 3.5 4B | 36.04 | #1175 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=hievidr-bench-wiki-rag · How It Works · Data refreshed daily, snapshot 2026-11-13.