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

Policy Following on POLICYSHIFTBENCH (Adaptive split)

89Accuracy

Human Performance

Updated 18d ago

Evaluation Results

Method	Links
Human Performance 2026.07		89	90.1	-
Human Performance 2026.07		88	89.1	-
PolicyShiftGuard-3B 2026.07		86.6	85.6	79
PolicyShiftGuard-7B 2026.07		86.3	86.8	73.8
PolicyShiftGuard-7B 2026.07		78.1	76.9	72.1
Gemini-3-Flash-Preview 2026.07		74.2	70.6	50.6
PolicyShiftGuard-3B 2026.07		73.8	66.7	64.8
Qwen3.5-35B-A3B 2026.07		71.3	63.5	36.2
Gemini-3-Flash-Preview 2026.07		70.1	67	45.4
GPT-5.4 2026.07		67.8	57.8	45.5
Qwen3.5-35B-A3B 2026.07		67.2	52.4	29.8
GPT-5.4 2026.07		66.6	57.8	42.7
Qwen3.5-2B 2026.07		66.4	18.2	0
Qwen2.5-VL-72B 2026.07		66.2	61.7	32.1
Qwen3.5-35B-A3B 2026.07		65.8	63.7	25.4
Qwen3.5-4B 2026.07		64.3	37.9	26.3
Qwen3.5-4B 2026.07		62.8	31.6	17.8
Qwen2.5-VL-72B 2026.07		62.1	49.4	27.4
Claude-Sonnet-4.6 2026.07		61.1	52.5	25.7
Qwen3.5-35B-A3B 2026.07		60.2	48.1	18.6
Qwen2.5-VL-32B 2026.07		59.9	59.2	16.1
Qwen3.5-2B 2026.07		57.7	9.1	0
Claude-Sonnet-4.6 2026.07		57.3	39.8	18.8
Qwen2.5-VL-32B 2026.07		56.5	47.2	9.4
SafeGuard-VL-RL-7B 2026.07		55.8	62.5	6.3
Qwen2.5-VL-7B 2026.07		54.5	29	9.5
SafeGuard-VL-RL-7B 2026.07		53.4	51	4
Qwen3.5-4B 2026.07		52.9	59	3
Qwen2.5-VL-7B 2026.07		52.6	20.6	4.8
Qwen3.5-4B 2026.07		52.3	47.4	4.4
GuardReasoner-VL-3B 2026.07		52.2	59.2	3.2
GuardReasoner-VL-3B 2026.07		52.1	63.6	4.2
Llama Guard-4-12B 2026.07		52	29.2	2
GuardReasoner-VL-7B 2026.07		51.6	61.5	10.8
GuardReasoner-VL-7B 2026.07		51.6	57.9	10.5
ShieldGemma2-4B 2026.07		51.4	26.8	4.8
Llama Guard-4-12B 2026.07		51.4	19.6	1.6
Qwen2.5-VL-3B 2026.07		50.9	61	2.4
ShieldGemma2-4B 2026.07		50.8	36.1	4.5
Qwen3.5-0.8B 2026.07		50.7	3.5	2.4
Qwen3.5-0.8B 2026.07		50.7	3.9	1.8
Qwen2.5-VL-3B 2026.07		50	65.9	2
QwenGuard-7B 2026.07		47.2	39.2	9.6
QwenGuard-7B 2026.07		44.6	39	10.2