Skill Adapter: Adaptive Skill Routing Boosts Agent Success Rates by 11%

Problem: Scaling Skill Libraries Creates Two Contradictions

When an agent's skill library grows from a few experiences to hundreds or thousands, two contradictions are amplified: how to adapt the same workflow to different users, and how to continuously select the right skills during long‑task execution.

The article illustrates this with a contract translation example (Figure 1): personalized requirements change the subsequent workflow, not just the output wording. General skills reuse the task flow but are not complete execution plans; user preferences directly alter skill combinations, demanding adaptivity.

Core Framework: Separate Skills from Adapters, Route by Execution State

Skill Adapter decomposes the system into three cooperating parts:

Separate general skills from Adapters. General skills store goals, steps, tool calls, and I/O constraints. Adapters record applicability conditions, local modifications, and feedback evidence, and can be added, updated, or deleted. The main flow stays stable while user differences are captured as reusable patches.

Online skill library supplies skills on demand. The main model first decomposes the task into a DAG. A routing agent then combines the current subtask, execution history, and environment feedback to select skill‑Adapter pairs. The main model reads only what the current node needs; subsequent nodes pre‑load candidates and verify/replace them once real input is known.

GRPO optimizes both decision types separately. The modification policy's reward comes from the change in held‑out task success rate. The routing reward considers both task success and skill selection F1. The main execution model remains frozen; training focuses on how to modify skills and how to select them.

Main Experiments: Higher Success Rates, Lower Execution Time

The paper compares Skill Adapter against ReAct, full‑skill, BM25, SkillRouter, Graph of Skills, and SkillRL on three benchmarks. Relative to SkillRL:

ALFWorld: success rate rises from 87.14% to 95.48% (+8.34 pp); average time drops from 70.6 s to 61.9 s .

WebShop: success rate rises from 72.73% to 81.53% (+8.80 pp); task score rises from 85.20 to 92.67 ; time falls to 50.7 s.

SkillsBench: success rate rises from 21.43% to 38.57% (+17.14 pp); average reward rises from 25.33 to 34.90 ; time drops from 529.4 s to 398.9 s.

Average success rate improvement across the three benchmarks is 11.43 percentage points . Notably, providing all skills at once performs poorly, showing that more context does not equal more appropriate skill supply .

Larger Skill Pools Demand Structured Routing

When the skill pool expands from 50 to 1,000, all methods degrade but at different rates: Skill Adapter falls from ~95.5% to 80.4% , while SkillRL drops from ~87.2% to 60.3% . At the 1,000‑skill scale Skill Adapter still leads by about 20 pp, demonstrating the combined value of structured retrieval, version filtering, and online refresh.

Ablation Study: Sources of Improvement

Adding the general skill pool lifts average success rate from 29.28% to 41.86% .

Adding skill combination relations and usage guidance reaches 48.35%; adding DAG pre‑loading reaches 52.75% .

Training only the modification policy yields 60.16%; training only the routing policy yields 65.28%; training both jointly reaches 71.86% .

The full method outperforms any single‑module configuration, confirming that personalized modification and online routing are complementary.

arXiv:2611.17007v1
https://arxiv.org/pdf/2611.17007