Commonsense Reasoning on CommonsenseQA Multiple Client (test)
80.61Client 1 AccuracyGRAD-TRANSFORMER
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GRAD-TRANSFORMERDescription=Learning to generate updates for LLMs2026.05 | 80.61 | 83.67 | 83.67 | 79.59 | 80.61 | 81.63 | 91.96 | |
| PTDescription=Ceiling performance, LLM fine-tuned directly on client data2026.05 | 80.61 | 86.73 | 81.63 | 81.63 | 79.59 | 82.04 | — | |
| W2SDescription=Weak-to-strong distillation2026.05 | 78.57 | 84.69 | 77.55 | 82.65 | 78.57 | 80.41 | 68.04 | |
| ConfDescription=Confidence-based weak-to-strong distillation2026.05 | 78.57 | 85.71 | 75.51 | 80.61 | 80.61 | 80.2 | 63.92 | |
| VisSupDescription=Weak-to-strong distillation baseline2026.05 | 78.57 | 85.71 | 76.53 | 81.63 | 79.59 | 80.41 | 68.04 | |
| PSDescription=TinyLM fine-tuned from the client side2026.05 | 77.55 | 78.57 | 73.47 | 76.53 | 78.57 | 76.94 | — |