SurveyReview: leaderboard
SurveyReview (2026) measures how closely automatic survey-paper evaluators match human peer reviewers over 675 surveys, scoring predicted review scores against annotations on a minus two to plus two scale. Mean squared error averaged over the four review dimensions, from 0 to 16; lower is better.
Metric: Mean squared error of predicted review scores, averaged over readability, criticalness, comprehensiveness and structure (0 to 16, lower is better; zero-shot LLM-as-a-judge with the unified dimension-specific rubric prompt on the full survey text, against scores annotated from real peer-review reports of the test-split surveys (MOPRD, F1000Research, OpenReview) on the {-2, -1, +1, +2} scale). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 2.28 | #172 |
| 2 | Claude Opus 4.5 (Thinking) | 2.77 | #152 (Claude Opus 4.5) |
| 3 | Qwen 3 32B | 3.21 | #960 |
| 4 | GLM-4.7 | 3.62 | #365 |
| 5 | Gemini 3 Pro (Preview) | 3.94 | #111 |
| 6 | DeepSeek V3.2 | 3.97 | #403 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=surveyreview · How It Works · Data refreshed daily, snapshot 2026-11-12.