Facet-Probe: leaderboard
Metric: Flip rate (% of items whose normalized answer changes across six sampled orderings of the five ordering facets (the mean of the five facet columns); lower is better). Source: arxiv.org. Saturation forecast: Around December 2026. 18 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 13 |
| 2 | Claude Opus 4.7 (Thinking) | 17 |
| 3 | Gemini 3 Flash (Preview) | 18 |
| 4 | GPT-5.5 (Medium) | 19 |
| 5 | Claude Sonnet 4.6 (Thinking) | 26 |
| 6 | GPT-5.4 Mini (Medium) | 31 |
| 7 | Qwen 3.5 9B | 36 |
| 8 | Qwen 3.5 4B | 40 |
| 9 | MedGemma-4B-IT | 49 |
| 10 | Qwen 3.5 2B | 59 |
| 11 | Qwen 3.5 0.8B | 62 |
Interactive version: theaggregate.ai/benchmark?slug=facet-probe · How It Works · Data refreshed daily, snapshot 2026-09-26.