Generalization Reasoning on GPQA diamond (Average@16 accuracy)
40.4Average@16 AccuracyHiLL
Evaluation results
| Method | Links | |
|---|---|---|
| HiLLBackbone=Qwen2.5-7B-Instruct2026.04 | 40.4 | |
| HiLLw/o TWBackbone=Qwen2.5-7B-Instruct2026.04 | 38.6 | |
| LUFFYBackbone=Qwen2.5-7B-Instruct2026.04 | 38.1 | |
| SAGEBackbone=Qwen2.5-7B-Instruct2026.04 | 38 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.04 | 37.2 | |
| BaseBackbone=Qwen2.5-7B-Instruct2026.04 | 37.1 | |
| Scaf-GRPOBackbone=Qwen2.5-7B-Instruct2026.04 | 36.6 | |
| HiLLw/o TWBackbone=Llama-3.2-3B-Instruct2026.04 | 28 | |
| HiLLBackbone=Llama-3.2-3B-Instruct2026.04 | 27.8 | |
| SAGEBackbone=Llama-3.2-3B-Instruct2026.04 | 27.3 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.04 | 26.3 | |
| Scaf-GRPOBackbone=Llama-3.2-3B-Instruct2026.04 | 24.1 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.04 | 17.9 | |
| LUFFYBackbone=Llama-3.2-3B-Instruct2026.04 | 16 |