The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents

ID: fe359ce0
原文链接: https://arxiv.org/abs/2607.22520
PDF: https://arxiv.org/pdf/2607.22520v1
作者: Darshan Tank, Baran Nama
日期: 2026-07-24
更新: 2026-07-24
分类: agents
来源类型: article
标签: llm-agents, skills, regression, evaluation, grounding, verification
质量评分: 5/5
抓取时间: 2026-07-29T06:19:39Z

中文导读

为LLM Agent添加程序性技能通常以平均任务成功率来评估，但这掩盖了一个关键代价：技能也可能导致性能倒退作者在约6000次实验运行中发现三种倒退机制：技能描述渗透（即使未被调用也改变Agent行为）接地偏移（程序步骤覆盖输入解释）验证偏移（程序抑制了Agent本应执行的输出检查）最佳技能的优势主要在于倒退更少，而非增益更大Agent可靠性更多依赖接地和验证环节，而非程序性技能的选择

为什么值得关注

为LLM Agent添加程序性技能通常以平均任务成功率来评估，但这掩盖了一个关键代价：技能也可能导致性能倒退作者在约6000次实验运行中发现三种倒退机制：技能描述渗透（即使未被调用也改变Agent行为）接地偏移（程序步骤覆盖输入解释）验证偏移（程序抑制了Agent本应执行的输出检...

本文以 arXiv 摘要与条目既有摘要为据：程序性技能/自进化/形式化研究/路由等主题直接对应 Agent 工程与评估实践中的可靠性、可验证性与部署权衡。

关键信息

论文标题：The Regression Tax: Decomposing Why Skills Help and Hurt LLM Agents
作者：Darshan Tank, Baran Nama
arXiv：https://arxiv.org/abs/2607.22520
发布时间：2026-07-24
更新时间：2026-07-24
arXiv 分类：cs.AI
关联标签：llm-agents, skills, regression, evaluation, grounding, verification

English Abstract

Adding procedural skills to an LLM agent is typically evaluated by average improvement in task success. However, this metric hides an important cost: skills can also make agents worse. We measure both sides by comparing agents with and without skills across nearly 6,000 runs spanning two office automation benchmarks and three model harness stacks. This allows us to distinguish two outcomes. A regression is a task solved without skills but failed after skills are added. A residual failure is a task that fails both with and without skills. We find that regressions are substantial enough that the best performing skills outperform others primarily by regressing less, not by gaining more. We identify three causes of regression: (i) skill description osmosis, a skill changes an agent's behavior simply by being present in context, even when it is never invoked; (ii) grounding displacement, a skill's prescribed procedure overrides how the agent interprets its inputs; and (iii) verification displacement, where the procedure suppresses checks the agent would otherwise perform on its outputs. Analysing persistent failures reveals the same underlying pattern. Existing skills overemphasize procedural guidance the stage least often responsible for failure while under supporting grounding and verification, the dominant sources of remaining errors. After correcting evaluation artifacts and studying traces, we find many regressions and persistent failures recoverable through better grounding and verification. Procedural skills should be evaluated by decomposing their net effect into gains and regressions, not by aggregate improvement alone. We identify three regression modes skills should avoid, and find that reliability depends more on grounding and verification than on procedural skill choice.

English Summary

Adding procedural skills to LLM agents is typically evaluated by average task improvement, but this hides a critical cost: skills can also cause regressions. Across ~6,000 runs on two office automation benchmarks and three model harness stacks, the authors identify three regression mechanisms: (1) skill description osmosis (skills alter agent behavior even when never invoked), (2) grounding displacement (procedures override input interpretation), and (3) verification displacement (procedures suppress output checks). The best-performing skills win primarily by regressing less, not by gaining more. Reliability depends more on grounding and verification than on procedural skill choice.

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。