Multiple Choice Question Answering on MMLU-Redux (test)
84.54AccuracyQwen2.5-32B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen2.5-32BScale=32B2025.10 | 84.54 | — | — | — | |
| GaC + SAFEEnsembling models=Qwen2.5 + EXAONE3.52025.10 | 83.79 | — | — | — | |
| Instruct Model (Q1)Backbone=Qwen3-8B2026.01 | 83.28 | 0.1337 | 0.125 | -0.6991 | |
| Ours (Label-Free)Backbone=Qwen3-8B, Protocol=Label-Free Post-hoc2026.01 | 83.28 | 0.1232 | 0.0659 | -0.7096 | |
| GaCEnsembling models=Qwen2.5 + EXAONE3.52025.10 | 82.73 | — | — | — | |
| Ours (Label-Free)Backbone=Ministral-3-8B, Protocol=Label-Free Post-hoc2026.01 | 80.52 | 0.1519 | 0.0664 | -0.6534 | |
| EXAONE3.5-32BScale=32B2025.10 | 79.26 | — | — | — | |
| Instruct Model (Q1)Backbone=Ministral-3-8B2026.01 | 79.21 | 0.172 | 0.1628 | -0.6201 | |
| Base Model (Q0)Backbone=Qwen3-8B2026.01 | 77.35 | 0.1229 | 0.0326 | -0.6506 | |
| Base Model (Q0)Backbone=Ministral-3-8B2026.01 | 75.63 | 0.1278 | 0.0534 | -0.6285 | |
| Instruct Model (Q1)Backbone=LLaMA-3.1-8B2026.01 | 71.74 | 0.245 | 0.2417 | -0.4724 | |
| Ours (Label-Free)Backbone=LLaMA-3.1-8B, Protocol=Label-Free Post-hoc2026.01 | 71.74 | 0.2153 | 0.1679 | -0.5022 | |
| Base Model (Q0)Backbone=LLaMA-3.1-8B2026.01 | 65.42 | 0.158 | 0.0176 | -0.4962 | |
| TS (Supervised)Backbone=Qwen3-8B, Protocol=Supervised Post-hoc2026.01 | — | 0.1116 | 0.0584 | -0.7212 | |
| TS (Supervised)Backbone=LLaMA-3.1-8B, Protocol=Supervised Post-hoc2026.01 | — | 0.1874 | 0.0352 | -0.5299 | |
| TS (Supervised)Backbone=Ministral-3-8B, Protocol=Supervised Post-hoc2026.01 | — | 0.1571 | 0.0546 | -0.635 |