Generalization Reasoning on Out-of-distribution Aggregate
51Average AccuracyHiLL
Evaluation Results
| Method | Links | |
|---|---|---|
| HiLLBackbone=Qwen2.5-7B-Instruct2026.04 | 51 | |
| HiLLw/o TWBackbone=Qwen2.5-7B-Instruct2026.04 | 49.2 | |
| LUFFYBackbone=Qwen2.5-7B-Instruct2026.04 | 48.6 | |
| SAGEBackbone=Qwen2.5-7B-Instruct2026.04 | 48.6 | |
| Scaf-GRPOBackbone=Qwen2.5-7B-Instruct2026.04 | 47.5 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.04 | 47.4 | |
| BaseBackbone=Qwen2.5-7B-Instruct2026.04 | 46.7 | |
| HiLLBackbone=Llama-3.2-3B-Instruct2026.04 | 35.3 | |
| HiLLw/o TWBackbone=Llama-3.2-3B-Instruct2026.04 | 34.5 | |
| SAGEBackbone=Llama-3.2-3B-Instruct2026.04 | 34 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.04 | 33.1 | |
| Scaf-GRPOBackbone=Llama-3.2-3B-Instruct2026.04 | 31 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.04 | 22.5 | |
| LUFFYBackbone=Llama-3.2-3B-Instruct2026.04 | 21.4 |