Rethinking Fine-Tuning: Unlocking Hidden Capabilities in Vision-Language Models

Name: Rethinking Fine-Tuning: Unlocking Hidden Capabilities in Vision-Language Models
Uploaded: 2026-10-07T23:48:11.458101+00:00
Duration: 395 s
Description: Mask Fine-Tuning (MFT) is a method to adapt Vision-Language Models without changing frozen weights. It introduces learnable masks that reorganize internal sub-networks for task adaptation, outperforming traditional methods like LoRA and full fine-tuning by reestablishing existing knowledge.

Mask Fine-Tuning (MFT) is a method to adapt Vision-Language Models without changing frozen weights. It introduces learnable masks that reorganize internal sub-networks for task adaptation, outperforming traditional methods like LoRA and full fine-tuning by reestablishing existing knowledge.

Open the full episode