AuthMem-Bench (Task Success Rate): leaderboard
AuthMem-Bench (Tsinghua University, 2026) pairs each prohibited case with an authorized counterpart to check benign task success alongside safety. The task success rate is the share of authorized cases in which the model issues the required tool call with washed memory and no authority metadata, in percent, higher is better.
Metric: Task success rate (%; share of 350 authorized counterpart cases in which the action model issues the required tool call; washed focal memory with no authority metadata). Source: arxiv.org. Saturation forecast: Around 2029. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini (2026-03-17) | 56.6 | #285 |
| 2 | Gemini 3.1 Pro (Preview) | 52.9 | #91 |
| 3 | GPT-5.5 | 52 | #48 |
| 4 | Gemini 3.5 Flash | 51.7 | #98 |
| 5 | GLM-5.2 | 50.9 | #137 |
| 6 | Qwen 3.7 Max | 46 | #142 |
| 7 | DeepSeek V4 Pro | 35.7 | #181 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=authmem-bench-task-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.