SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning

์ €์ž: Hanzhen Wang, Jiaming Xu, Yushun Xiang, Jiayi Pan, Yongkang Zhou, Yong-Lu Li, Guohao Dai | ๋‚ ์งœ: 2025-09-06 | URL: https://arxiv.org/abs/2509.05614 📄 PDF


Essence

Figure 2

Figure 2. Overview of SpecPrune-VLA. We prune the visual tokens with global and local information with a lightweight act

SpecPrune-VLA๋Š” Vision-Language-Action ๋ชจ๋ธ์˜ LLM ์ถ”๋ก ์„ ๊ฐ€์†ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์‹œ๊ฐ„-๊ณต๊ฐ„ ์ผ๊ด€์„ฑ์„ ํ™œ์šฉํ•œ ์•ก์…˜-์ธ์‹ ์ž์ฒด-์ถ”์ธก ํ† ํฐ ํ”„๋ฃจ๋‹ ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ๋‘ ๋‹จ๊ณ„ ํ”„๋ฃจ๋‹(์•ก์…˜ ๋ ˆ๋ฒจ ์ •์  ํ”„๋ฃจ๋‹๊ณผ ๋ ˆ์ด์–ด ๋ ˆ๋ฒจ ๋™์  ํ”„๋ฃจ๋‹)๊ณผ ์•ก์…˜-์ธ์‹ ์ปจํŠธ๋กค๋Ÿฌ๋ฅผ ํ†ตํ•ด ์ตœ๋Œ€ 1.70๋ฐฐ ์†๋„ ํ–ฅ์ƒ์„ ๋‹ฌ์„ฑํ•œ๋‹ค.

Motivation

Achievement

Figure 3

Figure 3. Insight 1: (a) Layers of different depth focus on different information. (b)(c)(d) In pick and place task, ran

How

Figure 2

Figure 2. Overview of SpecPrune-VLA. We prune the visual tokens with global and local information with a lightweight act

Originality

Limitation & Further Study

Evaluation

Novelty: 4/5 Technical Soundness: 3/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: SpecPrune-VLA๋Š” VLA ๋ชจ๋ธ์˜ spatial-temporal consistency๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  ์ด๋ฅผ ํ™œ์šฉํ•œ ์ƒˆ๋กœ์šด ํ”„๋ฃจ๋‹ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์—ฌ ์‹ค์งˆ์ ์ธ ์†๋„ ํ–ฅ์ƒ๊ณผ ์„ฑ๋Šฅ ์œ ์ง€๋ฅผ ๋™์‹œ์— ๋‹ฌ์„ฑํ–ˆ๋‹ค. Training-free ๋ฐฉ์‹์˜ ์ผ๋ฐ˜์„ฑ๊ณผ ๋ช…ํ™•ํ•œ ์‹คํ—˜ ๊ฒ€์ฆ์ด ๊ฐ•์ ์ด๋ฉฐ, VLA ๋ชจ๋ธ ์ตœ์ ํ™”์˜ ์ค‘์š”ํ•œ ์ง„์ „์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.

← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •