JudgeSense - Relevance: leaderboard
JudgeSense (2026) measures how often an LLM judge gives the same verdict when one judging request is worded two equivalent ways, at temperature 0, on items from human-labelled corpora. Agreement across the two wordings when picking the more relevant of two TREC-COVID passages, in both candidate orders, from 0 to 1; higher is better.
Metric: Judge Sensitivity Score: share of items on which the judge returns the same verdict under two reworded instruction templates that ask for the same judgement, over the pairs it answered (refusals and malformed outputs left out); temperature 0 (claude-opus-4-7 at the provider default), 1,024 output tokens, on 250 BEIR TREC-COVID topics, each with a fully relevant passage and a human-rejected one, judged in both candidate orders (500 rows); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 25 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.7 Flash | 0.982 | #41 |
| 2 | Kimi K3 | 0.972 | #61 |
| 3 | Claude Opus 4.7 | 0.964 | #85 |
| 4 | Llama 4 Maverick Instruct FP8 | 0.962 | #1023 |
| 5 | DeepSeek V4 Pro (0813) | 0.96 | #97 |
| 6 | Qwen 3.7 Flash (2026-07-15) | 0.954 | |
| 7 | Qwen 3 32B | 0.952 | #972 |
| 8 | GLM-5.2 | 0.95 | #150 |
| 9 | Qwen 3.8 27B | 0.946 | #221 |
| 10 | DeepSeek V4 Flash (0731) | 0.946 | #153 |
| 11 | Mistral Medium (2604) | 0.946 | #679 |
| 12 | Qwen 3.6 35B A3B | 0.944 | #481 |
| 13 | JudgeSense Novita deepseek-v4-flash (checkpoint unspecified) | 0.944 | |
| 14 | Qwen 2.5 72B Instruct | 0.94 | #1117 |
| 15 | Qwen 3 14B | 0.932 | #1269 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=judgesense-relevance · How It Works · Data refreshed daily, snapshot 2026-11-21.