Reasoning with Latent Activations on Rail2Country
93.7R2C Mono AccuracyAR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ARModel Backbone=Gemma2 9B2025.10 | 93.7 | 86 | |
| DeepSeek-R1-8BModel Backbone=DeepSeek-R1-8B2025.10 | 83.7 | 60 | |
| GPT-4o2025.10 | 82.7 | 69 | |
| ARModel Backbone=Llama3.1 8B2025.10 | 74.7 | 62.7 | |
| Llama3.1 70B itModel Backbone=Llama3.1 70B, Type=Instruction-tuned2025.10 | 68.3 | 33.3 | |
| Gemma2 27B itModel Backbone=Gemma2 27B, Type=Instruction-tuned2025.10 | 61 | 45 | |
| Llama3.1 8BModel Backbone=Llama3.1 8B2025.10 | 41 | 29.7 | |
| Gemma2 9BModel Backbone=Gemma2 9B2025.10 | 35.3 | 25.7 |