文章背景与核心概要
现代编码智能体（Coding Agents）通常通过长期运行的进程来执行任务，其中包含大量模型调用、工具交互和代码修改。为了在性能和成本之间取得平衡，一种常见的操作策略是动态切换模型：当任务变复杂时升级到高性能（HC）模型，而在完成复杂推理后降级到低性能（LC）模型以节省成本。
本文引入了“交接税”（Handoff Tax）这一概念，用于描述当一个模型必须接续由另一个模型生成的“非原生”轨迹（即动作和上下文序列）时所产生的性能与成本惩罚。作者通过研究证明，全轨迹升级只能弥补不到一半的模型层级质量差距，同时还会产生高昂的成本；相反，降级则能提供更具成本效益的权衡。研究进一步表明，处理这些交接的最佳方式取决于切换的方向：减少轨迹信息有助于提升升级质量，而保留轨迹信息对于维持降级期间的质量至关重要。
The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents
arXiv: 2608.24358
Authors: Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur, Ron Litman
Submitted: August 25, 2026
Subject: Artificial Intelligence (cs.AI)
Summary
现代编码智能体通常通过长期运行的进程来执行任务，其中包含大量的模型调用、工具交互和代码修改。一种常见的操作策略是动态切换模型——当任务变复杂时升级到高性能（HC）模型，或者在完成繁重的推理后降级到低性能（LC）模型以节省成本。
Modern coding agents often operate through long-running processes involving numerous model calls, tool interactions, and code modifications. A common operational strategy is to dynamically switch between models—escalating to high-capability (HC) models when tasks become complex, or downshifting to low-capability (LC) models to save costs once the heavy reasoning is complete.
本文引入了“交接税”（"Handoff Tax"）这一概念：当一个模型必须接续另一个模型生成的“非原生”轨迹（动作和上下文序列）时，所遭受的性能和成本惩罚。作者证明，全轨迹升级所恢复的质量差距不到模型层级差距的一半，同时却产生了巨额成本。相反，降级则提供了更高效的成本-质量权衡。该研究进一步表明，处理这些交接的最佳方式取决于切换的方向：减少轨迹信息可以提高升级质量，而保留轨迹信息对于在降级期间维持质量至关重要。
This paper introduces the concept of the "Handoff Tax": the performance and cost penalty incurred when one model must continue a "non-native" trajectory (the sequence of actions and context) generated by another model. The authors demonstrate that full-trajectory escalation recovers less than half of the quality gap between model tiers while incurring significant costs. Conversely, downshifting provides a more efficient cost-quality trade-off. The study further reveals that the optimal way to handle these handoffs depends on the direction of the switch: reducing trajectory information improves escalation quality, while retaining it is vital for maintaining quality during downshifts.
Key Findings
1. The Handoff Tax
研究人员发现，在不同能力水平的模型之间转移任务时，存在一种持续存在的惩罚。尽管成本更高，但从 LC 模型升级到 HC 模型未能完全弥补性能差距，这证实了“交接”本身引入了阻碍接收方性能的摩擦。
The researchers identified a consistent penalty when transferring tasks between models of different capability levels. Escalating from an LC model to an HC model fails to fully bridge the performance gap, despite the higher cost, confirming that the "handoff" itself introduces friction that hinders the receiver's performance.
2. Directional Asymmetry
该研究强调了轨迹信息应如何管理的四个关键反转： * 升级（LC 到 HC）： 减少从 LC 模型继承的轨迹信息量实际上会提高 HC 模型的性能。 * 降级（HC 到 LC）： 移除 HC 模型的轨迹信息会降低 LC 模型输出的质量。
The study highlights a critical reversal in how trajectory information should be managed: * Escalation (LC to HC): Reducing the amount of inherited trajectory information from the LC model actually improves the performance of the HC model. * Downshift (HC to LC): Removing the trajectory information from the HC model reduces the quality of the LC model’s output.
3. Practical Implications
研究结果表明，基于 LLM 的智能体开发者不应将模型交接视为“即插即用”的操作。相反，他们应该实施针对切换方向定制的上下文管理策略，以兼顾成本效率和任务成功率。
The findings suggest that developers of LLM-based agents should not treat model handoffs as a "plug-and-play" operation. Instead, they should implement context-management strategies tailored to the direction of the switch to optimize for both cost-efficiency and task success.
Accessing the Paper
Metadata: This paper is indexed under cs.AI. For further bibliographic tools, including BibTeX citations and related research, please visit the official arXiv abstract page.