PolicyShiftBench - Adaptive Split (PSS): leaderboard

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); Adaptive split, 1,000 instances over 130 images under 16 policies from the training policy families; mean of three evaluations). Source: arxiv.org. Saturation forecast: Around January 2027. 19 models tracked.

Top models

#ModelScore
1Gemini 3 Flash (Preview)45.4
2GPT-5.442.7
3Qwen 3.5 35B A3B (Thinking)36.2
4Qwen 3.5 4B (Thinking)26.3
5Claude Sonnet 4.625.7
6Qwen 3.5 35B A3B (Non-reasoning)25.4
7Qwen 2.5 VL 7B9.5
8Qwen 3.5 4B (Non-reasoning)3
9Qwen 3.5 0.8B (Non-reasoning)2.4
10Qwen 3.5 2B (Non-reasoning)0

Interactive version: theaggregate.ai/benchmark?slug=policyshiftbench-adaptive-split-pss · How It Works · Data refreshed daily, snapshot 2026-09-29.