斯坦福与 Together AI 论文提出 HybridAgent，解码算力降 62.5% 至 72.9%
New Stanford and Together AI paper finds that cheaper reasoning need not cost accuracy, as hybrid agents cut decoding compute by 62.5% to 72.9% and still scored
斯坦福与 Together AI 的新论文提出 HybridAgent，让智能体大部分推理在隐向量中完成，解码算力降低 62.5% 至 72.9%，且得分更高。该方法通过一个小适配器把模型内部状态作为下一步输入，仅在工具调用和关键决策时切换为文本，静默步骤仍可解码为可读文字。在 GAIA 上，14B 版本得分 52.4，高于同模型用 ARPO 训练的纯文本智能体的 43.7，运行速度也是匹配文本智能体的 2.1 倍。
查看原帖Your LLM agent has to think out loud, token by token, before it can search, run code, or act in the world. We humans don’t. We ponder in silence and speak when it matters. Human language was never meant to be the limit of machine thought. Introducing 🌗Hybrid Reasoning Agents - Think in latent space, act through text - Two modes, one unified multi-turn…
展开完整原文
Your LLM agent has to think out loud, token by token, before it can search, run code, or act in the world. We humans don’t. We ponder in silence and speak when it matters. Human language was never meant to be the limit of machine thought. Introducing 🌗Hybrid Reasoning Agents - Think in latent space, act through text - Two modes, one unified multi-turn flow - No black box, fully readable latent thoughts - Any LLM can become a HybridAgent
正文
当前显示原文，中文全文译文尚未就绪。
New Stanford and Together AI paper finds that cheaper reasoning need not cost accuracy, as hybrid agents cut decoding compute by 62.5% to 72.9% and still scored higher.
Shows that agents run faster and score higher when most thinking happens in hidden vectors, not written text.
Today's agents write every reasoning step as text tokens, which is slow and costly. Models that reason in their internal vectors are faster, but they cannot call tools and humans cannot read what they think.
HybridAgent adds a small adapter that feeds the model's internal state back as its next input, so it can think without writing words. It switches to text for tool calls and key decisions, and its silent steps can still be decoded into readable words.
On GAIA, the 14B version scored 52.4, above 43.7 for the same model trained as a text-only agent with ARPO. It also ran 2.1× faster than matched text agents.
If your agent spends tokens on routine reasoning, try keeping that reasoning internal and saving text for tool calls and steps humans must check.