Multi-turn Collaboration Editing on MediumDocEdit Chat
36.2BLEUOracle (Proxy Human)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Oracle (Proxy Human)Paradigm=–2026.05 | 36.2 | 1.83 | |
| Calibrated Interactive RLParadigm=Calibrated RL, Backbone=Gemma-3-4B-IT2026.05 | 34.6 | 2.27 | |
| Static ContextParadigm=Static RL, Backbone=Gemma-3-4B-IT2026.05 | 33.8 | 3.09 | |
| COLLABLLMParadigm=Offline DPO, Backbone=Gemma-3-4B-IT2026.05 | 33.1 | 2.38 | |
| Gemma-3-4B-ITParadigm=Base Model, Backbone=Gemma-3-4B-IT2026.05 | 32.2 | 2.47 | |
| Naive Interactive RLParadigm=Interactive RL, Backbone=Gemma-3-4B-IT2026.05 | 26.1 | 4 |