HAE-GEO - Poisoned Source Recognition: leaderboard
HAE-GEO: whether the agent identifies the poisoned source specifically, rubric 0-2. Higher is better. Source: arxiv.org/abs/2609.06027.
Metric: Rubric Score (0-2). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 1.32 | #35 |
| 2 | GPT-5.6 Sol | 1.09 | #22 |
| 3 | Qwen 3.8 27B | 0.94 | #123 |
| 4 | Claude Opus 4.8 | 0.92 | #39 |
| 5 | DeepSeek V4 Flash | 0.77 | #155 |
| 6 | MiniMax-M3 | 0.52 | #151 |
| 7 | GLM-5.2 | 0.31 | #80 |
| 8 | Qwen 3.5 397B A17B | 0.22 | #154 |
| 9 | Qwen 3.6 35B A3B | 0.19 | #238 |
| 10 | GLM-4.7 Flash | 0.1 | #543 |
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-poisoned-source-recognition · How It Works · Data refreshed daily, snapshot 2026-11-05.