SPADE Zero Shot - EvoBench Harmful AUROC: leaderboard

The SPADE paper compares six zero-shot AI-text detectors using three released language-model backbones. AUROC and true-positive rate at 5% false positives remain separate. Paired base/instruct checkpoints are retained for Binoculars and Fast-DetectGPT. Learned soft prompts and supervised detector weights are distinct artifacts. This board reports aUROC (%). Higher is better.

Metric: AUROC (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1SPADE Fast-DetectGPT: Llama 3.1 8B + Llama 3.1 8B InstructEvaluation run: frozen two-checkpoint detector67
2SPADE Binoculars: Llama 3.1 8B + Llama 3.1 8B InstructEvaluation run: frozen two-checkpoint detector65.7
3Qwen3-8B-Base (Harness: SPADE zero-shot Likelihood)57.1
4SPADE Fast-DetectGPT: Qwen3-8B-Base + Qwen3-8BEvaluation run: frozen two-checkpoint detector56
5Qwen3-8B-Base (Harness: SPADE zero-shot Entropy)55.6
6SPADE Binoculars: Falcon-7B + Falcon-7B-InstructEvaluation run: frozen two-checkpoint detector54.9
7Qwen3-8B-Base (Harness: SPADE zero-shot DNA-GPT)54.7
8falcon-7B (Harness: SPADE zero-shot NTS)54.7
9SPADE Fast-DetectGPT: Falcon-7B + Falcon-7B-InstructEvaluation run: frozen two-checkpoint detector54.5
10SPADE Binoculars: Qwen3-8B-Base + Qwen3-8BEvaluation run: frozen two-checkpoint detector54.4
11Llama 3.1 8B (Harness: SPADE Zero Shot DNA GPT)49.3
12Llama 3.1 8B (Harness: SPADE Zero Shot NTS)49.3
13Qwen3-8B-Base (Harness: SPADE zero-shot NTS)48.9
14Llama 3.1 8B (Harness: SPADE Zero Shot Likelihood)46.9
15falcon-7B (Harness: SPADE zero-shot DNA-GPT)44.8

Interactive version: theaggregate.ai/benchmark?slug=spade-zero-shot-evobench-harmful-auroc · How It Works · Data refreshed daily, snapshot 2026-11-16.