Researchers have developed a new technique called VETO to address shortcut learning in Video Large Language Models (VideoLLMs). This method suppresses the tendency of VideoLLMs to rely on textual regularities in training data, a shortcut that bypasses actual video understanding. VETO achieves this by projecting supervised fine-tuning updates onto the orthogonal complement of the text-only activation subspace, ensuring the model grounds its answers in the full video input. AI
IMPACT This method could improve the reliability and generalization of video understanding models by preventing them from exploiting superficial textual patterns.
RANK_REASON The cluster contains a research paper detailing a new method for improving VideoLLMs. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
- supervised fine-tuning
- VETO
- Video-Evidence Training via Orthogonal projection
- VideoLLM
- VideoLLMs
- Youngwoo-git
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →