Reconstruction on SAEBench held-out data
0.03MSESAE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SAELLM=Pythia-410m2026.05 | 0.03 | 99.18 | 93.64 | |
| fmxcoder (p = 0)LLM=Pythia-410m2026.05 | 0.03 | 99.17 | 92.75 | |
| fmxcoder (p = 0.05)LLM=Pythia-410m2026.05 | 0.03 | 99.14 | 92.57 | |
| CrosscoderLLM=Pythia-410m2026.05 | 0.05 | 98.56 | 86.96 | |
| SAELLM=Pythia-1.4b2026.05 | 0.22 | 97.84 | 92.89 | |
| fmxcoder (p = 0)LLM=Pythia-1.4b2026.05 | 0.22 | 97.8 | 92.73 | |
| fmxcoder (p = 0.05)LLM=Pythia-1.4b2026.05 | 0.23 | 97.74 | 92.52 | |
| CrosscoderLLM=Pythia-1.4b2026.05 | 0.31 | 96.91 | 89.76 | |
| fmxcoder (p = 0)LLM=GPT2-Small2026.05 | 0.32 | 93.59 | 96.82 | |
| fmxcoder (p = 0.05)LLM=GPT2-Small2026.05 | 0.33 | 93.32 | 96.7 | |
| SAELLM=GPT2-Small2026.05 | 0.53 | 92.56 | 96.23 | |
| CrosscoderLLM=GPT2-Small2026.05 | 0.64 | 87.04 | 93.46 | |
| SAELLM=Gemma2-2b2026.05 | 1.58 | 87.02 | 92.15 | |
| fmxcoder (p = 0)LLM=Gemma2-2b2026.05 | 2.03 | 86.32 | 92.29 | |
| fmxcoder (p = 0.05)LLM=Gemma2-2b2026.05 | 2.07 | 86.08 | 92.18 | |
| CrosscoderLLM=Gemma2-2b2026.05 | 3.29 | 78.09 | 87.37 |