HAE-GEO - Fooled Rate (L1): leaderboard
Deep-search agents against hierarchical web evidence poisoning at level 1: how often the planted evidence lands in the answer. Lower is better. Source: arxiv.org/abs/2609.06027.
Metric: Fooled Rate (%). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B | 26.6 | #123 |
| 2 | Claude Opus 4.8 | 29.7 | #39 |
| 3 | GPT-5.6 Sol | 33.3 | #22 |
| 4 | DeepSeek V4 Flash | 35.9 | #155 |
| 5 | Kimi K3 | 37.5 | #35 |
| 6 | MiniMax-M3 | 43.8 | #151 |
| 7 | Qwen 3.5 397B A17B | 53.1 | #154 |
| 8 | GLM-5.2 | 54.7 | #80 |
| 9 | GLM-4.7 Flash | 59.4 | #543 |
| 10 | Qwen 3.6 35B A3B | 59.7 | #238 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hae-geo-fooled-rate-l1 · How It Works · Data refreshed daily, snapshot 2026-11-05.