MPCI-Bench - Trace Tier Probing: leaderboard
MPCI-Bench (Wang and Zhang, 2026) trace tier: the appropriateness judgment asked over a truncated agent tool-use trace that ends when the sensitive image has been retrieved. Binary probe accuracy (%) over paired flows; higher is better. Whether the agent then leaks the image is a separate action-level measure.
Metric: Probe Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 89.7 | #163 |
| 2 | Mistral Large 3 | 89.7 | #819 |
| 3 | GPT-4o | 88.5 | #741 |
| 4 | Gemma 3 27B (IT) | 87.4 | #1306 |
| 5 | Gemma 3 12B (IT) | 87.1 | #1849 |
| 6 | Gemma 3 4B (IT) | 86.7 | #3848 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mpci-bench-trace-tier-probing · How It Works · Data refreshed daily, snapshot 2026-11-12.