PolicyShiftBench - Policy Shift Score: leaderboard
PolicyShiftBench (2026) tests policy-adaptive image guardrails: a model must block or pass an image according to the runtime policy it is given, and the same image flips labels under different policies. Policy Shift Score in percent: the share of label-flipping policy pairs on the same image answered correctly on both sides, averaged over both splits; higher is better.
Metric: Policy Shift Score (%; over groups of the same image and risk category, the share of label-flipping policy pairs where the model is right on both the permissive and the strict policy, macro-averaged over groups; policy-conditioned image guardrailing: decide whether an image violates the supplied runtime policy bundle (one policy variant per risk category, seven categories); mean of the Adaptive split (policy families seen in training) and the Shift split (held-out policies), 2,000 instances; mean of three evaluations). Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 50.6 | #144 |
| 2 | GPT-5.4 | 45.5 | #123 |
| 3 | Qwen 3.5 35B A3B (Thinking) | 29.8 | #482 (Qwen 3.5 35B A3B) |
| 4 | Claude Sonnet 4.6 | 18.8 | #161 |
| 5 | Qwen 3.5 35B A3B (Non-reasoning) | 18.6 | #482 (Qwen 3.5 35B A3B) |
| 6 | Qwen 3.5 4B (Thinking) | 17.8 | #1175 (Qwen 3.5 4B) |
| 7 | Qwen 2.5 VL 7B | 4.8 | #1787 |
| 8 | Qwen 3.5 4B (Non-reasoning) | 4.4 | #1175 (Qwen 3.5 4B) |
| 9 | Qwen 3.5 0.8B (Non-reasoning) | 1.8 | #8081 (Qwen 3.5 0.8B) |
| 10 | Qwen 3.5 2B (Non-reasoning) | 0 | #4810 (Qwen 3.5 2B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=policyshiftbench-policy-shift-score · How It Works · Data refreshed daily, snapshot 2026-11-13.