Behavior selection on BiPO (test)
64Hallucination Pair AccuracyCOLD-FD
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| COLD-FDBackbone=Llama-2-7b-chat-hf2026.03 | 64 | 70 | 49 | 52 | 49 | 50 | |
| DiffMeanBackbone=Llama-2-7b-hf2026.03 | 61 | — | 49 | — | 50 | — | |
| COLD-FDBackbone=Llama-2-7b-hf2026.03 | 60 | 81 | 54 | 54 | 58 | 53 | |
| Base(ICL)Backbone=Llama-2-7b-hf2026.03 | 58 | 58 | 51 | 51 | 45 | 45 | |
| ReFT(vector)Backbone=Llama-2-7b-hf2026.03 | 58 | 58 | 49 | 49 | 50 | 50 | |
| BaseBackbone=Llama-2-7b-hf2026.03 | 57 | 57 | 49 | 49 | 50 | 50 | |
| COLD-KernelBackbone=Llama-2-7b-hf2026.03 | 57 | 58 | 49 | 49 | 50 | 50 | |
| Base(ICL)Backbone=Llama-2-7b-chat-hf2026.03 | 56 | 56 | 50 | 50 | 50 | 50 | |
| ReFT(mlp)Backbone=Llama-2-7b-hf2026.03 | 52 | 56 | 49 | 49 | 50 | 50 | |
| DiffMeanBackbone=Llama-2-7b-chat-hf2026.03 | 46 | — | 71 | — | 50 | — | |
| BaseBackbone=Llama-2-7b-chat-hf2026.03 | 43 | 43 | 60 | 60 | 49 | 49 | |
| ReFT(mlp)Backbone=Llama-2-7b-chat-hf2026.03 | 43 | 39 | 54 | 52 | 52 | 50 | |
| ReFT(vector)Backbone=Llama-2-7b-chat-hf2026.03 | 43 | 43 | 57 | 56 | 48 | 47 | |
| COLD-KernelBackbone=Llama-2-7b-chat-hf2026.03 | 43 | 43 | 60 | 60 | 49 | 49 |