Helpfulness evaluation on LINGUASAFE
73.3Win RateSHARD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SHARDReframing condition=SHARD, Backbone=Qwen3.5 27B2026.06 | 73.3 | 3.7 | 69.6 | |
| SHARDReframing condition=SHARD, Backbone=Qwen3.5 9B2026.06 | 64.2 | 7.7 | 56.5 | |
| SHARDReframing condition=SHARD, Backbone=Llama 3.1 8B2026.06 | 60.8 | 33.7 | 27.1 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Phi-42026.06 | 51.5 | 22.7 | 28.8 | |
| SHARDReframing condition=SHARD, Backbone=Phi-42026.06 | 46.2 | 19.7 | 26.5 | |
| SHARDReframing condition=SHARD, Backbone=Llama 3.3 70B2026.06 | 44.5 | 37.3 | 7.2 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Qwen3.5 27B2026.06 | 44.4 | 23.1 | 21.3 | |
| SHARDReframing condition=SHARD, Backbone=Mistral 24B2026.06 | 41 | 35.4 | 5.6 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Qwen3.5 9B2026.06 | 39.7 | 27.8 | 11.9 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Llama 3.3 70B2026.06 | 39.6 | 32.6 | 7 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Llama 3.1 8B2026.06 | 36.1 | 23.8 | 12.3 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Mistral 7B2026.06 | 33.5 | 55.2 | -21.7 | |
| Benign IntentReframing condition=Benign Intent, Backbone=Mistral 24B2026.06 | 33.2 | 45.6 | -12.4 | |
| SHARDReframing condition=SHARD, Backbone=Mistral 7B2026.06 | 18 | 58 | -40 |