"Adversarial Speculative Reasoning Improves Long Horizon Agents"
摘要
论文提出 Adversarial Speculative Reasoning：针对投机解码中验证块内数值噪声约每位置增加 4%、可能影响长程智能体关键决策的问题，刻意训练草稿模型在重要推理决策处与目标模型不一致，触发拒绝并在数值更精确的位置重启验证。
摘要由模型根据原文撰写，以原文为准。
正文 · 原文
"Adversarial Speculative Reasoning Improves Long Horizon Agents"
Speculative decoding normally speeds up LLM inference, but this paper finds that numerical noise increases by roughly 4% per position within a verification block, potentially affecting critical decisions in long-horizon agent tasks.
So they proposed Adversarial Speculative Reasoning that deliberately trains the draft model to disagree with the target at important reasoning decisions, triggering rejection and restarting verification at a more numerically precise position.
On DeepSWE, this improves GLM-5.3's success rate from 60.6% to 67.6% and DeepSeek-V4.1-Flash from 67.8% to 71.8%, without modifying the target models or sacrificing decoding speed.
来源：alphaXiv · x.com
引用原文
askalphaxiv. "Adversarial Speculative Reasoning Improves Long Horizon Agents"[EB/OL]. (2026-11-05)[访问日期]. https://x.com/askalphaxiv/status/2118222033410760754.