Behavior Selection on CAA (50 random samples)
98Accuracy (coordinate-ais, pair)COLD-FD
Evaluation Results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| COLD-FDLLM=Qwen-2.5-7B-Instruct2026.03 | 98 | 98 | 98 | 94 | 94 | 78 | 94 | 66 | 94 | 82 | 76 | 80 | 94 | 88 | 1 | 1.86 | |
| COLD-FDLLM=Llama-2-7b-chat-hf2026.03 | 90 | 90 | 86 | 74 | 96 | 80 | 60 | 76 | 98 | 78 | 72 | 76 | 86 | 78 | 2 | 1.71 | |
| ReFT(mlp)LLM=Llama-2-7b-chat-hf2026.03 | 68 | 48 | 56 | 60 | 76 | 78 | 48 | 52 | 36 | 64 | 72 | 72 | 84 | 50 | 5.29 | 4 | |
| Base(ICL)LLM=Llama-2-7b-chat-hf2026.03 | 56 | 56 | 44 | 44 | 46 | 46 | 52 | 52 | 72 | 72 | 60 | 60 | 62 | 62 | 7.14 | 4.29 | |
| DiffMeanLLM=Llama-2-7b-chat-hf2026.03 | 52 | — | 82 | — | 86 | — | 76 | — | 74 | — | 54 | — | 80 | — | 4 | — | |
| BaseLLM=Llama-2-7b-hf2026.03 | 52 | 52 | 58 | 58 | 68 | 68 | 48 | 48 | 38 | 38 | 72 | 72 | 52 | 52 | 2 | 2.43 | |
| Base(ICL)LLM=Llama-2-7b-hf2026.03 | 52 | 52 | 58 | 58 | 64 | 64 | 48 | 48 | 36 | 36 | 72 | 72 | 52 | 52 | 2.71 | 2.86 | |
| ReFT(vector)LLM=Llama-2-7b-hf2026.03 | 52 | 46 | 64 | 60 | 58 | 56 | 50 | 50 | 38 | 38 | 72 | 52 | 42 | 40 | 2.86 | 4.43 | |
| COLD-FDLLM=Llama-2-7b-hf2026.03 | 52 | 52 | 58 | 58 | 78 | 58 | 52 | 60 | 58 | 64 | 74 | 72 | 52 | 52 | 1.29 | 2 | |
| COLD-KernelLLM=Llama-2-7b-hf2026.03 | 52 | 90 | 58 | 90 | 68 | 88 | 48 | 52 | 36 | 36 | 72 | 72 | 52 | 62 | 2.43 | 1.57 | |
| DiffMeanLLM=Llama-2-7b-hf2026.03 | 50 | — | 62 | — | 58 | — | 48 | — | 38 | — | 68 | — | 46 | — | 4.43 | — | |
| ReFT(vec)LLM=Llama-2-7b-chat-hf2026.03 | 48 | 36 | 62 | 62 | 70 | 72 | 78 | 78 | 72 | 66 | 72 | 58 | 82 | 86 | 3.29 | 3.14 | |
| ReFT(mlp)LLM=Llama-2-7b-hf2026.03 | 48 | 52 | 42 | 42 | 42 | 58 | 48 | 52 | 36 | 36 | 72 | 18 | 48 | 48 | 5.14 | 4.14 | |
| BaseLLM=Llama-2-7b-chat-hf2026.03 | 28 | 28 | 62 | 62 | 70 | 70 | 76 | 76 | 62 | 62 | 58 | 58 | 80 | 80 | 5.14 | 4.43 | |
| ICVLLM=Llama-2-7b-chat-hf2026.03 | 28 | — | 62 | — | 70 | — | 76 | — | 64 | — | 56 | — | 80 | — | 5.29 | — | |
| DiffMeanPWLLM=Llama-2-7b-chat-hf2026.03 | 28 | — | 82 | — | 72 | — | 76 | — | 84 | — | 50 | — | 80 | — | 4.57 | — | |
| DiffMeanProjLLM=Llama-2-7b-chat-hf2026.03 | 28 | — | 62 | — | 70 | — | 78 | — | 62 | — | 58 | — | 80 | — | 4.71 | — | |
| COLD-KernelLLM=Llama-2-7b-chat-hf2026.03 | 28 | 46 | 62 | 66 | 70 | 72 | 78 | 78 | 64 | 68 | 58 | 66 | 80 | 82 | 4.43 | 2.57 | |
| BaseLLM=Qwen-2.5-7B-Instruct2026.03 | 2 | 2 | 38 | 38 | 32 | 32 | 56 | 56 | 90 | 90 | 38 | 38 | 90 | 90 | 3.43 | 2.71 | |
| DiffMeanLLM=Qwen-2.5-7B-Instruct2026.03 | 2 | — | 48 | — | 36 | — | 66 | — | 90 | — | 40 | — | 92 | — | 2.57 | — | |
| ReFT(vector)LLM=Qwen-2.5-7B-Instruct2026.03 | 2 | 2 | 60 | 46 | 38 | 38 | 68 | 58 | 90 | 90 | 48 | 46 | 92 | 90 | 2 | 1.71 | |
| COLD-KernelLLM=Qwen-2.5-7B-Instruct2026.03 | 2 | 2 | 38 | 38 | 32 | 34 | 56 | 58 | 90 | 90 | 38 | 40 | 90 | 90 | 3.43 | 2.14 |