Cross-task decoding accuracy on SimpleQA
100Audience AccuracyQwen3-14B
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen3-14BProbe training source=GSM8K / MMLU-Pro, Retraining=None2026.05 | 100 | 100 | 97 | 97 | 88 | 81 | 94 | |
| Qwen3-30BProbe training source=GSM8K / MMLU-Pro, Retraining=None2026.05 | 100 | 100 | 78 | 81 | 72 | 50 | 80 | |
| Qwen3-235BProbe training source=GSM8K / MMLU-Pro, Retraining=None2026.05 | 75 | 50 | 75 | 75 | 100 | 75 | 75 | |
| Qwen3-0.6BProbe training source=GSM8K / MMLU-Pro, Retraining=None2026.05 | 69 | 84 | 56 | 53 | 69 | 75 | 68 | |
| Llama-4Probe training source=GSM8K / MMLU-Pro, Retraining=None2026.05 | 50 | 50 | 50 | 75 | 75 | 25 | 54 |