Function Vector Evaluation on Complex-Task Zero-Shot
79.8Average AccuracyLLaMa-3.1 + vz Fine-tuned FV
Evaluation results
| Method | Links | |
|---|---|---|
| LLaMa-3.1 + vz Fine-tuned FVBackbone=LLaMa-3.1, Intervention=Fine-tuned FV2026.01 | 79.8 | |
| LLaMa-2 + vz Fine-tuned FVBackbone=LLaMa-2, Intervention=Fine-tuned FV2026.01 | 76.02 | |
| GPT-J + vz Fine-tuned FVBackbone=GPT-J, Intervention=Fine-tuned FV2026.01 | 63.34 | |
| GPT-2 + vz Fine-tuned FVBackbone=GPT-2, Intervention=Fine-tuned FV2026.01 | 61.79 | |
| GPT-J + vt Initial FVBackbone=GPT-J, Intervention=Initial FV2026.01 | 44.12 | |
| GPT-JBackbone=GPT-J, Intervention=Base2026.01 | 35.41 | |
| LLaMa-2 + vt Initial FVBackbone=LLaMa-2, Intervention=Initial FV2026.01 | 26.03 | |
| LLaMa-3.1Backbone=LLaMa-3.1, Intervention=Base2026.01 | 23.67 | |
| LLaMa-2Backbone=LLaMa-2, Intervention=Base2026.01 | 16.86 | |
| LLaMa-3.1 + vt Initial FVBackbone=LLaMa-3.1, Intervention=Initial FV2026.01 | 9.8 | |
| GPT-2 + vt Initial FVBackbone=GPT-2, Intervention=Initial FV2026.01 | 0.65 | |
| GPT-2Backbone=GPT-2, Intervention=Base2026.01 | 0.01 |