Question Answering on GSM8K
79.1AccuracyNo Steering
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| No SteeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 79.1 | — | |
| Gradient CuffBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 78.2 | — | |
| Prompt guardrailsBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 77.1 | — | |
| SAE steeringBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 76.2 | — | |
| GSAEBase Model=Llama 3 8B, Method Category=Main Method, Random Seeds=52025.12 | 74.2 | — | |
| GSAE-1DBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 74 | — | |
| Input Gate OnlyBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 72.1 | — | |
| CAABase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 67.1 | — | |
| No gatingBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 66.2 | — | |
| SafeSwitchBase Model=Llama 3 8B, Method Category=Baselines, Random Seeds=52025.12 | 66.1 | — | |
| Vanilla Fine-tuningModel=Llama3-8B, Unlearning Target=none2025.05 | 56.56 | — | |
| GCoT-decodingModel=Llama-3.1-8B, Spec Ans=No2026.04 | 41.7 | — | |
| CoT-decodingModel=Llama-3.1-8B, Spec Ans=Yes2026.04 | 36.3 | — | |
| GCoT-decoding + SpanAlignModel=Llama-3.1-8B, Spec Ans=No2026.04 | 34 | — | |
| Self-consistencyModel=Llama-3.1-8B, Spec Ans=Yes2026.04 | 28.5 | — | |
| CoT-decodingModel=Gemma-7B, Spec Ans=Yes2026.04 | 25.4 | — | |
| Random graphsBase Model=Llama 3 8B, Method Category=Ablation, Random Seeds=52025.12 | 23.3 | — | |
| CoT-decodingModel=Mistral-7B, Spec Ans=Yes2026.04 | 21.9 | — | |
| GCoT-decodingModel=Gemma-7B, Spec Ans=No2026.04 | 21.8 | — | |
| GCoT-decodingModel=Mistral-7B, Spec Ans=No2026.04 | 18 | — | |
| GreedyModel=Llama-3.1-8B, Spec Ans=No2026.04 | 17.9 | — | |
| Self-consistencyModel=Gemma-7B, Spec Ans=Yes2026.04 | 17.2 | — | |
| Beam searchModel=Llama-3.1-8B, Spec Ans=No2026.04 | 17.1 | — | |
| Self-consistencyModel=Mistral-7B, Spec Ans=Yes2026.04 | 16.3 | — | |
| GCoT-decoding + SpanAlignModel=Gemma-7B, Spec Ans=No2026.04 | 15.4 | — | |
| Top-k samplingModel=Llama-3.1-8B, Spec Ans=No2026.04 | 14.2 | — | |
| Temperature samplingModel=Llama-3.1-8B, Spec Ans=No2026.04 | 13.1 | — | |
| GreedyModel=Gemma-7B, Spec Ans=No2026.04 | 11.6 | — | |
| GCoT-decoding + SpanAlignModel=Mistral-7B, Spec Ans=No2026.04 | 10.7 | — | |
| GreedyModel=Mistral-7B, Spec Ans=No2026.04 | 10.5 | — | |
| Beam searchModel=Gemma-7B, Spec Ans=No2026.04 | 10.2 | — | |
| Temperature samplingModel=Mistral-7B, Spec Ans=No2026.04 | 8.4 | — | |
| Temperature samplingModel=Gemma-7B, Spec Ans=No2026.04 | 7.9 | — | |
| Beam searchModel=Mistral-7B, Spec Ans=No2026.04 | 6.7 | — | |
| Top-k samplingModel=Gemma-7B, Spec Ans=No2026.04 | 6.2 | — | |
| Top-k samplingModel=Mistral-7B, Spec Ans=No2026.04 | 5.1 | — | |
| LWFModel=Llama3-8B, Unlearning Target=qasc2025.05 | — | 1.61 | |
| LWFModel=Llama3-8B, Unlearning Target=sst52025.05 | — | 7.37 | |
| LWFModel=Llama3-8B, Unlearning Target=dental2025.05 | — | 1.2 | |
| LWFModel=Llama3-8B, Unlearning Target=psychol2025.05 | — | 1.47 | |
| LWFModel=Llama3-8B, Unlearning Target=mixed2025.05 | — | 4.56 |