Autointerpretation on IOI
76AccuracySAE features
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SAE featuresSignal type=SAE features, Fuzzing protocol=AutoInterp2026.02 | 76 | — | — | |
| ACC++Backbone=GPT-2 Small, Signal type=ACC++ signals, Fuzzing protocol=fuzzing-style LLM-judge2026.02 | 70 | 0.68 | 0.85 | |
| ACC++Backbone=Pythia-160M, Signal type=ACC++ signals, Fuzzing protocol=fuzzing-style LLM-judge2026.02 | 65 | 0.62 | 0.85 | |
| ACC++Backbone=Gemma-2-2B, Signal type=ACC++ signals, Fuzzing protocol=fuzzing-style LLM-judge2026.02 | 60 | 0.58 | 0.8 |