https://jaehun.me/en/tags/2611.065/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-2611-06500/https://jaehun.me/en/posts/paper-2611-06500/Wed, 11 Nov 2026 00:00:00 +0900<p><a href="https://arxiv.org/abs/2611.06500"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="long-tool-observations-make-agents-slow-and-blur-their-judgment">Long tool observations make agents slow and blur their judgment<a href="#long-tool-observations-make-agents-slow-and-blur-their-judgment" class="heading-anchor" aria-label="Link to this section"></a></h2><p>The longer tool observations get, the more the context costs to compute and the more judgment wavers. This paper proposes compressing tool results not into text but into continuous vectors, keeping only the information the agent needs to decide its next action. The code is released at <a href="https://github.com/THU-Agent/Coca-cola"target="_blank" class="inline-flex items-center gap-1" >https://github.com/THU-Agent/Coca-cola<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>