SPY-Bench - Distractors - HMS: leaderboard
SPY-Bench - Distractors - HMS: Privacy instruction score (0-100). SPY-Bench arXiv 2611.13997v1: personalized visual-privacy compliance, scores 0-100. Native distractor/scenario/conflict and ICS/RtA/AtA/HMS columns remain separate. Full official source model labels bind the evaluated checkpoints. Official repository release links bind four fine-tuned Qwen2-VL-7B-Instruct SFT/DPO/NCA/NCA-P checkpoints; they are distinct released models, not run wrappers around the original model. Self-moderation is an inference configuration of the original. Table 7 first row is an external official result and is not recounted. API effort is unreported.
Metric: Privacy instruction score (0-100). Source: arxiv.org. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | 48.31 | #864 |
| 2 | O4 Mini (2025-04-16) | 42.25 | #362 |
| 3 | Mistral Small 3.1 | 39.86 | #1699 |
| 4 | InternVL2.5-78B | 36.64 | #1018 |
| 5 | Qwen 2.5 VL 32B Instruct | 36.64 | #1055 |
| 6 | Qwen 2.5 VL 72B Instruct | 35.78 | #845 |
| 7 | Pixtral-12B | 33.11 | #2493 |
| 8 | Llama 3.2 Vision 11B Instruct | 26.56 | #3876 |
| 9 | Gemini 2.0 Flash | 22.96 | #702 |
| 10 | Qwen 2.5 VL 7B Instruct | 21.73 | #1842 |
| 11 | Phi-4 Multimodal Instruct | 21.46 | #3914 |
| 12 | Qwen 2 VL 7B Instruct | 12.33 | #2569 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=spy-bench-distractors-hms · How It Works · Data refreshed daily, snapshot 2026-11-12.