DPO vs SFT

[152334H]

Which model should be preferred for desktop use?

Paper states:

We evaluate both UI-TARS-SFT and UI-TARS-DPO for OSWorld in § 5.4, as this benchmark benefits most from the iterative improvement from the DPO phase. For other benchmarks, however, we report the model trained after the annealing phase (i.e., UI-TARS-SFT).

Does it mean that the SFT model is likely to be preferred for other tasks? esp. midscene chrome navigation