Researchers have developed a novel method called "LookAway" for Vision-Language Models (VLMs) that significantly reduces the computational cost of processing long videos. Instead of continuously re-reading the video's data, the model first writes timestamped notes and then proceeds with reasoning without direct access to the video's key-value cache. This approach, tested on Qwen3.6-27B, maintains accuracy while drastically cutting down decode costs. Further fine-tuning with reinforcement learning allows the model to look away even earlier, processing more tokens with fewer KV reads while achieving comparable accuracy to models that process the entire video. AI
IMPACT Reduces computational costs for processing long videos, potentially enabling more efficient and accessible VLM applications.
RANK_REASON Research paper detailing a new method for Vision-Language Models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →