EarlyTom: Early Token Compression Completes Fast Video Understanding

Hesong Wang, Xin Jin, Lu Lu, Chenhaowen Li, Jian Chen, Qiang Liu, Huan Wang

arXiv:2605.30010 · 2026-05-29 공개 · arXiv · PDF

training-free token-compression vision-encoder throughput ttft video-llm early-stage llava-onevision

Abstract

Video large language models (Video-LLMs) have demonstrated strong capabilities in video understanding tasks. However, their practical deployment is still hindered by the inefficiency introduced by processing massive amounts of visual tokens. Although recent approaches achieve extremely low token retention ratios while maintaining accuracy comparable to full-token baselines, most of them perform compression only at the late stage of prefilling, leaving the efficiency of the vision encoder unoptimized. In this paper, we first show that vision encoding contributes a large portion to the time-to-first-token (TTFT). Therefore, instead of compressing visual tokens only after the vision encoder, performing compression inside the encoder still leaves substantial room for exploration. Based on this insight, we propose EarlyTom, a training-free token compression framework that performs early-stage visual token compression inside the vision encoder, enabling significantly better TTFT reduction and higher throughput. In addition, we introduce a decoupled spatial token selection strategy that improves the overall compression effectiveness. EarlyTom reduces TTFT by up to 2.65x and FLOPs by up to 61% on a single NVIDIA A100 GPU for the LLaVA-OneVision-7B model, while maintaining accuracy comparable to the full-token baseline. These improvements substantially enhance the practicality of deploying Video-LLMs in real-world production scenarios.

한국어 요약

📋 한 줄 요약

**[Video-LLM / Token Compression]** EarlyTom이 vision encoder 내부에서 early-stage visual token 압축 — TTFT 2.65×·FLOPs 61% 절감, decoupled spatial token selection으로 LLaVA-OneVision-7B에서 full-token 정확도 유지.

🎯 핵심 기여도

💡 핵심 아이디어

Video-LLM 효율의 핵심 병목은 vision encoder 자체에 있으며, encoder 후가 아닌 내부에서 early-stage compression을 수행하면 TTFT를 본질적으로 단축할 수 있고, decoupled spatial token selection이 압축률·정확도 균형을 더 잘 잡는다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Video-LLM 효율 병목의 정확한 위치(encoder 자체) 식별, encoder 내부 early-stage 압축이라는 새 카테고리, training-free·plug-in으로 즉시 채택 가능, TTFT/FLOPs 동시 감소로 production 가치. **한계**: LLaVA-OneVision 중심으로 다른 비디오 LLM 일반화 추가 검증, 매우 긴 비디오·복잡 쿼리에서의 압축 한계, decoupled selection의 hyperparameter 튜닝.

🚀 실용적 활용