Skip to main content
Back to timeline
arXivSource publication:

Skill Adapter hangs per-user adapters off shared agent skills, beating SkillRL by 11.43 points on average across three benchmarks

Synopsis

When LLM agents call external skills, the general procedure is shared but each user's tool choices, parameters and output requirements differ. Skill Adapter distills user-specific local adjustments into Adapters stored separately from general-purpose skills, while a routing agent selects skills and applicable Adapters online from the current subtask and execution state and incrementally maintains a task-level skill library. On ALFWorld, WebShop and SkillsBench, success rates rise by 8.34, 8.80 and 17.14 percentage points over SkillRL, and mean execution time falls on all three benchmarks.

Source-provided article image: Skill Adapter: Personalized Modification and Online Routing of General-Purpose Skills
arXiv

Interpretation

When agents use external skills, a general-purpose skill describes only the shared execution procedure for a class of tasks, while specific tool choices, operating practices and output requirements vary by user, and in long-horizon tasks subgoals, available inputs and environment states change as execution proceeds. Earlier skill pools organized experience into reusable skills to cut repeated exploration, but user-specific operating requirements had no place separate from the general procedure, and skill selection relied mainly on semantic matching between task and skill. Skill Adapter distills user requirements and execution feedback into Adapters attached to general-purpose skills; an Adapter records applicability conditions and local modifications to tools, parameters, steps and output checks, and one general skill can carry adapted versions for multiple users.

At execution time a routing agent picks general-purpose skills and their applicable Adapters from the current subtask, execution progress and environment feedback, and the main model reads the entries needed for the current node on demand. Routing no longer fixes skills once; as the task advances it reuses entries that remain valid, removes or replaces invalid ones, and preloads candidates for later nodes. The main model decomposes the task into a directed acyclic graph; routing retrieves candidates by topic search and semantic matching, expands along skill composition relations up to 30 candidates, and rechecks applicability once actual inputs are known, with modification and routing policies trained separately by GRPO while the main model stays frozen.

On ALFWorld, WebShop and SkillsBench, Skill Adapter leads all seven compared configurations on the core metrics. Against SkillRL, success rates rise from 87.14%, 72.73% and 21.43% to 95.48%, 81.53% and 38.57% on the three benchmarks. Mean execution time falls from 70.6 s, 57.8 s and 529.4 s to 61.9 s, 50.7 s and 398.9 s; with the skill pool grown to 1,000 skills, success drops from about 95.5% to 80.4% but stays about 20 points above SkillRL's roughly 60.3%.

Perspective

Evaluation covers six long-horizon task types in ALFWorld, product-purchase tasks in WebShop and external-skill use in SkillsBench, with test sets of 420, 1,500 and 210 tasks and interaction budgets of 50, 15 and 30 steps; Qwen3-8B executes the first two benchmarks and GPT-5-mini executes SkillsBench. The scaling experiment starts from 50 skills and adds distractor skills from other task categories up to 1,000.

Success rates fall for all methods as the skill pool grows, with Skill Adapter dropping from about 95.5% to 80.4%; personalization conditions are drawn from goals, operational constraints and output requirements explicitly stated in task instructions, and how preferences that generalize across tasks are used is described only in principle.