Question Answering on ARC-C (Accuracy, Avg Tokens)
0.88AccuracyDeepSeek-R1
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DeepSeek-R1Setting=ARM only2026.03 | 0.88 | 504 | |
| Qwen3-1.7BSetting=ARM only2026.03 | 0.86 | 397 | |
| LLaDA-8B → Llama-3.2-3BSetting=text-space2026.03 | 0.825 | 4 | |
| Latent-DARM (LLaDA-8B → Llama-3.2-3B)Setting=latent-space, Plan tokens=2562026.03 | 0.81 | 2 | |
| Llama-3.2-3BSetting=ARM only2026.03 | 0.795 | 4 | |
| Latent-DARM (LLaDA-8B → Llama-3.2-3B)Setting=latent-space, Plan tokens=642026.03 | 0.785 | 2 | |
| Latent-DARM (LLaDA-8B → Llama-3.2-3B)Setting=latent-space, Plan tokens=1282026.03 | 0.765 | 2 |