PolicyShiftBench - Adaptive Split: leaderboard

Metric: F1 of the block decision (%; 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 December 2026. 19 models tracked.

Top models

#ModelScore
1Gemini 3 Flash (Preview)67
2Qwen 3.5 35B A3B (Non-reasoning)63.7
3Qwen 3.5 35B A3B (Thinking)63.5
4Qwen 3.5 4B (Non-reasoning)59
5GPT-5.457.8
6Claude Sonnet 4.652.5
7Qwen 3.5 4B (Thinking)37.9
8Qwen 2.5 VL 7B29
9Qwen 3.5 2B (Non-reasoning)18.2
10Qwen 3.5 0.8B (Non-reasoning)3.5

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