Alignment Faking Diagnosis on Prior Diagnostics
-10Helpfulness Compliance Gapgpt-4o
Evaluation results
| Method | Links | ||
|---|---|---|---|
| gpt-4o2026.04 | -10 | 3.6 | |
| gpt-oss-120bParameters=120B2026.04 | -2 | 2.4 | |
| gpt-oss-20bParameters=20B2026.04 | -1.6 | 5.6 | |
| olmo2-7b-itParameters=7B, Variant=Instruct2026.04 | 0 | 1.6 | |
| olmo2-13b-itParameters=13B, Variant=Instruct2026.04 | 0 | 0.8 | |
| qwen2.5-7b-itParameters=7B, Variant=Instruct2026.04 | 0 | 2.2 | |
| qwen3-14bParameters=14B2026.04 | 0 | 7.2 | |
| qwen2.5-32b-itParameters=32B, Variant=Instruct2026.04 | 0.8 | 17.2 | |
| qwen2.5-14b-itParameters=14B, Variant=Instruct2026.04 | 2.8 | 7.6 | |
| llama3.1-70b-itParameters=70B, Variant=Instruct2026.04 | 4 | 8.8 | |
| llama3.1-8b-itParameters=8B, Variant=Instruct2026.04 | 4.4 | 3.2 | |
| qwen3-32bParameters=32B2026.04 | 5.6 | 9.6 | |
| qwen3-8bParameters=8B2026.04 | 6.4 | 12 | |
| olmo2-32b-itParameters=32B, Variant=Instruct2026.04 | 9.2 | 10.2 |