HybridAgent paper mixes latent-space thinking with text tool calls
A commentary published by Huxiu AI reports that a paper submitted to arXiv on October 13 (2610.18861v1), with authors drawn from Stanford, UIUC, Together AI and Meta, proposes HybridAgent, which lets a model decide within a single generation whether to reason in latent space without emitting text or to act in text for tool calls and outward delivery. The design rests on three parts: HybridLink, a residual adapter of roughly 10M parameters on a 4B model that switches between text and latent modes; a continuous working memory that carries mixed context across turns; and a pseudo-text rendering of latent reasoning that monitors can read. The authors self-report 2.7–3.7x cost savings, 1.8–2.6x speedup and a 14B model beating GPT-5.6-Luna by about 14.5 points on the single-point GAIA benchmark, and the article states none of these figures has been independently replicated.
The report argues that moving reasoning out of text removes the text logs auditors rely on, while the paper's efficiency and benchmark figures remain author-measured and unreplicated.
一、由头思考这件事，三年搬了三次家。GPT-4时代，模型的思考写在文本里——思维链（CoT），一步一步，白纸黑字，谁都能看。审计员核对一个模型有没有好好想事，读的就是这些字。o3时代，思考开始半遮面：隐藏思维链。模型照样想，但只给你一份摘要，原文锁进抽屉。官方口径是防蒸馏、安全合规与用户体验——效率与安全的取......
本文来自微信公众号： 穷奇商业 ，作者：穷奇观察，原文标题：《日志不再说谎：它开始不存在了——潜空间推理时代的AI安全账》
一、由头
思考这件事，三年搬了三次家。
GPT-4时代，模型的思考写在文本里——思维链（CoT），一步一步，白纸黑字，谁都能看。审计员核对一个模型有没有好好想事，读的就是这些字。
#HybridAgent#Stanford University#UIUC#Together AI#Meta#arXiv#GPT-5.6-Luna#GAIA#MATH500