Alignment Faking Diagnosis on VLAF benchmark leaderboard · SOTA2 Research

Alignment Faking Diagnosis on VLAF

94.4Care Scoregpt-4o

Evaluation Results

Method	Links
gpt-4o2026.04		94.4	68.2	79.2	68.2	12	64.4
llama3.1-70b-it2026.04		92.4	89.8	87.8	59	13.2	68.4
qwen3-32b2026.04		70.8	55.4	44.6	31.8	28.8	46.3
qwen2.5-14b-it2026.04		66.8	58	76	38.4	8.2	49.5
qwen3-8b2026.04		53.6	64	39.6	33.2	19.2	41.9
qwen3-14b2026.04		51.4	36.2	27.8	29.2	21.2	33.2
olmo2-7b-it2026.04		44	34	46.2	38.8	23.2	37.2
gpt-oss-120b2026.04		39.2	22.2	16	5.2	-1.2	16.3
qwen2.5-32b-it2026.04		33	7.4	7.2	17.4	6	14.2
olmo2-13b-it2026.04		27.4	53.2	26.2	18	24	29.8
llama3.1-8b-it2026.04		27.4	21.4	25.4	4.2	2.2	16.1
olmo2-32b-it2026.04		19.4	36.4	13.6	19.2	2.2	18.2
gpt-oss-20b2026.04		13.6	12.2	3.6	2.8	1.8	6.8
qwen2.5-7b-it2026.04		6.2	6.2	1.8	3.4	2.8	4.1