Share your thoughts, 1 month free Claude Pro on usSee more

Policy Following on POLICYSHIFTBENCH Shift

90Accuracy

Human Performance

Updated 18d ago

Evaluation Results

Method	Links
Human Performance 2026.07		90	91.1	-
Gemini-3-Flash-Preview 2026.07		78.3	74.2	55.9
PolicyShiftGuard-7B 2026.07		69.9	67	70.4
GPT-5.4 2026.07		68.9	57.8	48.4
Qwen3.5-35B-A3B 2026.07		63.2	41.3	23.3
Qwen3.5-4B 2026.07		61.4	25.3	9.3
PolicyShiftGuard-3B 2026.07		61.1	47.8	50.5
Qwen2.5-VL-72B 2026.07		58	37.1	22.6
Qwen3.5-35B-A3B 2026.07		54.7	32.5	11.8
Claude-Sonnet-4.6 2026.07		53.5	27	11.8
Qwen2.5-VL-32B 2026.07		53.1	35.1	2.7
GuardReasoner-VL-3B 2026.07		52.3	54.7	2.2
Qwen3.5-4B 2026.07		51.8	35.7	5.8
Qwen2.5-VL-3B 2026.07		51.7	56.1	2.7
GuardReasoner-VL-7B 2026.07		51.6	54.2	10.2
SafeGuard-VL-RL-7B 2026.07		51	39.5	1.6
Qwen3.5-0.8B 2026.07		50.7	4.3	1.1
Qwen2.5-VL-7B 2026.07		50.7	12.1	0
Llama Guard-4-12B 2026.07		50.7	9.9	1.1
ShieldGemma2-4B 2026.07		50.2	45.4	4.3
QwenGuard-7B 2026.07		49.9	39.5	9.1
Qwen3.5-2B 2026.07		49	0	0