Physical Reasoning on PIQA (PPL, Acc)
3.28PerplexityBase
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BaseBackbone=GPT-J2026.04 | 3.28 | 49.13 | |
| Selective Parameter OptimizationBackbone=GPT-J2026.04 | 3.38 | 50.49 | |
| LoRA(ν + μ)Backbone=GPT-J2026.04 | 3.45 | 50.44 | |
| EWCLoRABackbone=GPT-J2026.04 | 3.47 | 49.4 | |
| LoRA(μ)Backbone=GPT-J2026.04 | 3.54 | 50.16 | |
| RSLoRABackbone=GPT-J2026.04 | 3.65 | 49.62 | |
| Selective Parameter OptimizationBackbone=LLaMA-32026.04 | 4.92 | 51.16 | |
| BaseBackbone=LLaMA-32026.04 | 4.94 | 48.53 | |
| EWCLoRABackbone=LLaMA-32026.04 | 4.98 | 48.45 | |
| LoRA(μ)Backbone=LLaMA-32026.04 | 5.43 | 48.75 | |
| LoRA(ν + μ)Backbone=LLaMA-32026.04 | 5.58 | 46.91 | |
| RSLoRABackbone=LLaMA-32026.04 | 5.71 | 47.5 |