RUTA method drastically cuts visual tokens for LLMs while preserving performance

Researchers have developed RUTA, a novel method for reducing the number of visual tokens processed by large language models. RUTA learns to select and allocate tokens based on query-specific relevance and a rate-utility objective, balancing task performance with computational cost. In evaluations, RUTA significantly reduced token usage on models like LLaVA-NeXT-7B and Qwen3-VL-8B while retaining a high percentage of task performance. AI

IMPACT This method could significantly reduce the computational cost of using vision-language models with high-resolution images and long videos.

RANK_REASON Publication of a new research paper on arXiv detailing a novel method for visual token allocation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• LLaVA-NeXT-7B
• Qwen3 VL 8B
• RUTA
• ScienceCast

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →