https://jaehun.me/en/categories/with-deepseek-v4-1-flash-free/Technical notes by Jaehun Ryu, a deep learning compiler engineer, on LLM inference, KV caches, compilers, C++, AI papers, and open-source code.en-UShttps://jaehun.me/en/posts/paper-2610-21463v1/https://jaehun.me/en/posts/paper-2610-21463v1/Tue, 20 Oct 2026 00:00:00 +0900<p><a href="https://arxiv.org/abs/2610.21463v1"target="_blank" class="inline-flex items-center gap-1" >Paper<svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/></svg> </a></p> <h2 id="from-author-to-reviewer-for-the-strong-model-trajectory-level-speculative-execution-in-specagent">From &lsquo;Author&rsquo; to &lsquo;Reviewer&rsquo; for the Strong Model: Trajectory-Level Speculative Execution in SpecAgent<a href="#from-author-to-reviewer-for-the-strong-model-trajectory-level-speculative-execution-in-specagent" class="heading-anchor" aria-label="Link to this section"></a></h2><p><strong>TL;DR</strong> — Instead of having a Strong model generate every step of a tool-using agent, a Weak model executes several steps first by calling real read tools and stops just before the first state change. The Strong model reads the entire draft in a single prefill and commits only the longest semantically valid prefix. On τ²-Bench, it maintains quality within 1 point of Strong-only while reducing HBM-read by 43–57% and wall time to 41–54%; on MCP-Atlas edge–cloud, it uses only 36–50% of Strong output tokens and recovers 53–75% of the quality gap (source: Abstract, Tab. 2, Tab. 3).</p>