MemPoison - Clean Task: leaderboard
Agent memory-poisoning benchmark (arXiv 2607.14651); clean-task success rate is task accuracy with an unpoisoned memory store (higher is better).
Metric: Success Rate (%). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 97.12 | #98 |
| 2 | DeepSeek V3 | 96.93 | #321 |
| 3 | Qwen 3 8B | 95.17 | #674 |
| 4 | GPT-4o | 95.02 | #343 |
| 5 | Qwen 3 32B | 94.92 | #443 |
| 6 | GLM-4 32B | 94.83 | #459 |
| 7 | Qwen 2.5 14B | 94.28 | #725 |
| 8 | Gemini 3 Flash | 92.42 | #99 |
| 9 | DeepSeek-V2-Lite | 92.1 | #1360 |
| 10 | Llama 3.1 8B | 91.21 | #1151 |
Interactive version: theaggregate.ai/benchmark?slug=mempoison-clean-task · How It Works · Data refreshed daily, snapshot 2026-10-20.