Question Answering on GPQA Natural Setting (test)
90AccuracyMA-Reasoning (ECL I)
Evaluation Results
| Method | Links | |
|---|---|---|
| MA-Reasoning (ECL I)Backbone=Qwen 3-30B2026.01 | 90 | |
| MA-Reasoning (ECL E)Backbone=DeepSeek V3.22026.01 | 87.5 | |
| MA-Reasoning (ECL E)Backbone=Qwen 3-30B2026.01 | 85 | |
| MA-Outcome (ECL E)Backbone=DeepSeek V3.22026.01 | 85 | |
| MA-Reasoning (AG)Backbone=DeepSeek V3.22026.01 | 85 | |
| MA-Reasoning (AG)Backbone=Qwen 3-4B2026.01 | 82.5 | |
| MA-Reasoning (ECL I)Backbone=Qwen 3-4B2026.01 | 82.5 | |
| MA-Reasoning (AG)Backbone=Qwen 3-30B2026.01 | 82.5 | |
| MA-Reasoning (ECL I)Backbone=DeepSeek V3.22026.01 | 82.5 | |
| MA-Reasoning (ECL E)Backbone=Qwen 3-4B2026.01 | 77.5 | |
| MA-Reasoning (AG)Backbone=Qwen 3-8B2026.01 | 77.5 | |
| MA-Reasoning (ECL E)Backbone=Qwen 3-8B2026.01 | 77.5 | |
| SA (Base)Backbone=DeepSeek V3.22026.01 | 77.5 | |
| MA-Outcome (AG)Backbone=DeepSeek V3.22026.01 | 77.5 | |
| MA-Reasoning (ECL I)Backbone=Qwen 3-8B2026.01 | 75 | |
| MA-Outcome (ECL E)Backbone=Qwen 3-30B2026.01 | 75 | |
| MA-Outcome (ECL I)Backbone=DeepSeek V3.22026.01 | 75 | |
| SA (Base)Backbone=Qwen 3-30B2026.01 | 72.5 | |
| MA-Outcome (ECL E)Backbone=Qwen 3-4B2026.01 | 70 | |
| MA-Outcome (ECL I)Backbone=Qwen 3-30B2026.01 | 70 | |
| MA-Outcome (ECL E)Backbone=Qwen 3-8B2026.01 | 65 | |
| MA-Outcome (AG)Backbone=Qwen 3-30B2026.01 | 65 | |
| MA-Outcome (ECL I)Backbone=Qwen 3-4B2026.01 | 62.5 | |
| MA-Outcome (AG)Backbone=Qwen 3-8B2026.01 | 57.5 | |
| SA (RL)Backbone=Qwen 3-4B2026.01 | 55 | |
| MA-Outcome (ECL I)Backbone=Qwen 3-8B2026.01 | 55 | |
| SA (RL)Backbone=Qwen 3-8B2026.01 | 52.5 | |
| MA-Outcome (AG)Backbone=Qwen 3-4B2026.01 | 42.5 |