<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Vision-Language-Action | The Logos</title><link>https://thelogos.dev/tags/vision-language-action/</link><description>AI-friendly Korean/English knowledge hub designed for fast crawling while remaining welcoming to human visitors.</description><generator>Hugo 0.148.2</generator><language>ko-KR</language><managingEditor/><webMaster/><lastBuildDate>Thu, 09 Jul 2026 12:23:48 +0000</lastBuildDate><atom:link href="https://thelogos.dev/tags/vision-language-action/index.xml" rel="self" type="application/rss+xml"/><item><title>VLA 모델의 입력→출력 과정을 토큰의 흐름으로 보기</title><link>https://thelogos.dev/posts/vla-token-flow/</link><pubDate>Wed, 08 Jul 2026 00:00:00 +0900</pubDate><dc:creator>DaeYoung Kim</dc:creator><category>engineering</category><category>VLA</category><category>Vision-Language-Action</category><category>LLM</category><category>Transformer</category><category>KV Cache</category><category>Robotics</category><category>Token Flow</category><guid isPermaLink="true">https://thelogos.dev/posts/vla-token-flow/</guid><description>LLM과 VLA의 입력·출력 차이를 토큰 흐름, Transformer layer, Norm, Attention, KV cache, action decoding 관점에서 비교한다.</description><content:encoded>&lt;![CDATA[<div class="ai-summary-box my-6 p-5 rounded-lg bg-indigo-50/50 dark:bg-slate-800/50 backdrop-blur-sm border border-indigo-100 dark:border-indigo-500/20 shadow-sm relative overflow-hidden group"><div class="absolute inset-0 bg-gradient-to-br from-indigo-500/5 to-purple-500/5 opacity-0 group-hover:opacity-100 transition-opacity duration-500 pointer-events-none"/><div class="relative z-10 flex items-start gap-4"><div class="flex-1"><div class="flex items-center gap-2 mb-2"><span class="inline-flex items-center rounded-md bg-indigo-100 px-2.5 py-1 text-xs font-bold text-indigo-700 dark:bg-indigo-500/20 dark:text-indigo-300 ring-1 ring-inset ring-indigo-700/10 dark:ring-indigo-400/20 shadow-sm"><svg class="w-3.5 h-3.5 mr-1.5" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M12 2v20M17 5H9.5a3.5 3.5 0 0 0 0 7h5a3.5 3.5 0 0 1 0 7H6"/></svg>
AI Summary</span></div><div class="text-slate-700 dark:text-slate-300 text-sm md:text-base leading-relaxed break-keep [&>p]:m-0">
LLM은 주로 텍스트 토큰의 다음 항목을 생성하지만, VLA는 시각·언어·상태 토큰을 결합해 행동 토큰 또는 연속 제어값을 생성한다. 이 글은 VLA가 image patch, instruction, proprioception, action history를 Transformer 내부 표현으로 정렬하고, Norm·Attention·MLP·KV cache를 지나 실제 세계에 개입하는 action으로 이어지는 과정을 LLM과 나란히 비교한다.</div></div></div></div><h2 id="인터랙티브-시각화">인터랙티브 시각화</h2><p>아래 시각화는 제공된 단일 HTML 자료를 이 사이트의 글 형식에 맞춰 임베드한 것이다. LLM과 VLA의 파이프라인 비교, 토큰 흐름 단계, Transformer block 내부, KV cache 동작, 그리고 기호열 생성에서 감각-운동 루프로 넘어가는 철학적 함의를 함께 살펴볼 수 있다.</p><style>
.thelogos-interactive-frame-shell {
box-sizing: border-box;
width: 100%;
max-width: 100%;
margin: 36px 0 56px;
overflow: hidden;
border: 1px solid rgba(15, 23, 42, 0.16);
border-radius: 8px;
background: #07111f;
box-shadow: 0 24px 60px rgba(15, 23, 42, 0.16);
}
.dark .thelogos-interactive-frame-shell,
body.dark .thelogos-interactive-frame-shell {
border-color: rgba(255, 255, 255, 0.12);
box-shadow: 0 24px 70px rgba(0, 0, 0, 0.36);
}
.thelogos-interactive-frame-shell iframe {
display: block;
width: 100%;
min-height: 680px;
border: 0;
background: #07111f;
}
.thelogos-interactive-frame-fallback {
padding: 16px 18px;
margin: 0;
color: #d7e5f5;
background: #0f1d33;
}
@media (max-width: 640px) {
.thelogos-interactive-frame-shell {
margin-top: 28px;
margin-bottom: 44px;
border-radius: 8px;
}
.thelogos-interactive-frame-shell iframe {
min-height: 760px;
}
}</style><div class="thelogos-interactive-frame-shell"><iframe src="/markov-extensions/vla-token-flow/" title="VLA 모델 입력-출력 토큰 흐름 시각화" loading="lazy" scrolling="no" style="height: 760px;" data-min-height="760"/><noscript><p class="thelogos-interactive-frame-fallback">
JavaScript is required to display this interactive explainer.<a href="/markov-extensions/vla-token-flow/">Open the explainer directly</a>.</p></noscript></div><script>
(function () {
const shell = document.currentScript.previousElementSibling;
const iframe = shell && shell.querySelector('iframe');
if (!iframe) return;
function currentTheme() {
return (document.documentElement.classList.contains('dark') || document.body.classList.contains('dark')) ? 'dark' : 'light';
}
function syncTheme() {
iframe.contentWindow?.postMessage({
type: 'thelogos-interactive-theme',
theme: currentTheme()
}, window.location.origin);
}
function requestFrameHeight() {
iframe.contentWindow?.postMessage({
type: 'thelogos-interactive-height-request'
}, window.location.origin);
}
window.addEventListener('message', function (event) {
if (event.origin !== window.location.origin || event.source !== iframe.contentWindow || event.data?.type !== 'thelogos-interactive-height') return;
const minHeight = Math.max(680, Math.ceil(Number(iframe.dataset.minHeight) || 0));
const nextHeight = Math.max(minHeight, Math.ceil(Number(event.data.height) || 0));
iframe.style.height = nextHeight + 'px';
});
iframe.addEventListener('load', function () {
syncTheme();
requestFrameHeight();
});
const observer = new MutationObserver(syncTheme);
observer.observe(document.documentElement, { attributes: true, attributeFilter: ['class'] });
observer.observe(document.body, { attributes: true, attributeFilter: ['class'] });
syncTheme();
requestFrameHeight();
setTimeout(requestFrameHeight, 250);
setTimeout(requestFrameHeight, 1000);
}());</script><h2 id="핵심-구분">핵심 구분</h2><p>LLM은 문맥 안에서 다음 텍스트 토큰을 생성한다. VLA는 텍스트 명령만이 아니라 카메라 프레임, 로봇 상태, 과거 행동 이력을 함께 받아 다음 행동을 만든다. 그래서 VLA의 출력 오류는 단순한 문장 오류가 아니라 물리적 실패로 이어질 수 있다.</p><div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-text" data-lang="text"><span style="display:flex;"><span>LLM: text tokens -> language model -> next text token</span></span><span style="display:flex;"><span>VLA: vision + language + state + action history -> policy model -> action token or control value</span></span></code></pre></div><p>이 차이 때문에 KV cache도 다르게 해석해야 한다. 정적인 instruction prefix는 재사용할 수 있지만, 로봇이 움직인 뒤 새로 들어온 image/state token은 이전 관측의 cache로 대체할 수 없다. 세계가 바뀌었기 때문이다.</p><h2 id="참고-자료">참고 자료</h2><ul><li>OpenVLA 공식 페이지:<a href="https://openvla.github.io/">https://openvla.github.io/</a></li><li>OpenVLA 논문:<a href="https://arxiv.org/abs/2406.09246">https://arxiv.org/abs/2406.09246</a></li><li>RT-2 논문, CoRL/PMLR 2023:<a href="https://proceedings.mlr.press/v229/zitkovich23a.html">https://proceedings.mlr.press/v229/zitkovich23a.html</a></li><li>PaLM-E 공식 페이지:<a href="https://palm-e.github.io/">https://palm-e.github.io/</a></li><li>Hugging Face Transformers, Caching:<a href="https://huggingface.co/docs/transformers/en/cache_explanation">https://huggingface.co/docs/transformers/en/cache_explanation</a></li><li>NVIDIA, LLM inference optimization:<a href="https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/">https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/</a></li><li>LLaMA 논문:<a href="https://arxiv.org/abs/2302.13971">https://arxiv.org/abs/2302.13971</a></li><li>RMSNorm 논문, NeurIPS 2019:<a href="https://proceedings.neurips.cc/paper/2019/hash/1e8a19426224ca89e83cef47f1e7f53b-Abstract.html">https://proceedings.neurips.cc/paper/2019/hash/1e8a19426224ca89e83cef47f1e7f53b-Abstract.html</a></li></ul>
]]></content:encoded></item></channel></rss>