Skill Adapter Boosts LLM Agents with Personalized Skill Tweaks and Real‑Time Routing

What is Skill Adapter and why it matters?

A team from Nanjing University and Jiangnan University introduced Skill Adapter, a system that lets large language model (LLM) agents adjust shared skills to individual user needs and pick the right skill at each step of a long‑running task. The paper, posted on arXiv on 13 Nov 2026, argues that while LLM agents can invoke external tools, they often rely on a static pool of general‑purpose skills that do not capture user‑specific preferences such as layout preservation or output formatting. Without personalization, agents may waste time trying unsuitable tools or need extra post‑processing, hurting both speed and reliability.

How does personalized modification work?

Skill Adapter stores Adapters alongside each general‑purpose skill. An Adapter records a small set of adjustments – for example, a different parameter setting, an extra validation step, or a custom output check – that aligns the skill with a particular user’s constraints. When a user submits a request, the system builds a personalization context $P_u$ that aggregates explicit preferences and any feedback gathered during execution. The modification policy examines $P_u$ and the current execution trace $h_t$ to decide which Adapter (if any) should be attached to the chosen skill. This decision is rewarded during training by the improvement in task success rate after the modification is applied.

How does online routing select skills during long tasks?

Long‑horizon tasks consist of many subgoals, and the set of usable skills can shift as inputs, environment states, and intermediate results evolve. Skill Adapter introduces a routing agent that, at each step $t$, receives the combined context $x_{u,t}=\langle q, P_u, h_t\rangle$, where $q$ is the original query. The agent first performs semantic matching to retrieve a shortlist of candidate general‑purpose skills. It then checks the applicability of each skill’s Adapters against the live context – for instance, whether a PDF‑extraction tool can access the required file format or whether a layout‑reconstruction step has the necessary layout metadata.

If a candidate passes these checks, it is placed in a task‑level online skill library. The library is incrementally updated: as new inputs arrive, previously stored entries are re‑validated, and fresh candidates are pre‑loaded for upcoming nodes in the task graph. This dynamic selection reduces redundant retrieval work and keeps the agent’s plan coherent across the whole execution.

What performance gains were reported?

The authors evaluated Skill Adapter on three public benchmarks – ALFWorld, WebShop, and SkillsBench – and compared it against the prior state‑of‑the‑art system SkillRL. Across all three datasets, Skill Adapter lifted average task success rates by 11.43 percentage points and cut mean execution time, indicating more efficient planning and fewer dead‑end attempts. A scaling experiment demonstrated that even when the skill pool grew to 1,000 entries, the framework retained its advantage, suggesting that the online routing mechanism handles larger libraries without a proportional slowdown.

How can Grafy users experiment with Skill Adapter concepts?

While Grafy does not yet ship a built‑in Skill Adapter module, creators, developers, and researchers can explore the underlying ideas using existing tools:

• Grafy Reader (/reader) lets you upload the arXiv PDF, ask questions about the methodology, and extract the formal definitions of $P_u$, $h_t$, and the routing algorithm.
• Grafy Graph (/graph) can be used to model the skill DAG that the routing agent builds, visualizing how adapters attach to base skills and how the online library evolves during execution.
• For developers interested in prototyping the reinforcement‑learning loop, Grafy Code (/code) provides an AI‑assisted coding environment where you can implement a simple GRPO‑based policy and test it on a toy skill set.

By combining these apps, a Grafy user could recreate a miniature version of the Skill Adapter pipeline: define a few generic tool wrappers, write adapters as JSON patches, and let a small LLM orchestrate them while the routing logic runs in a Python script inside the Code studio.

Frequently asked questions

Does Skill Adapter change the underlying general‑purpose skills?

No. The framework keeps the original skill unchanged and layers user‑specific adjustments in separate Adapter objects, preserving the shared implementation for reuse.

How does the routing agent know when to refresh its skill library?

After each subtask, the agent re‑evaluates the stored entries against the updated context $x_{u,t}$; if inputs or environmental feedback no longer satisfy an Adapter’s conditions, the entry is discarded and new candidates are fetched.

Is the reported 11.43‑point boost specific to any particular benchmark?

The improvement is an average across the three benchmarks mentioned—ALFWorld, WebShop, and SkillsBench—each representing different domains (embodied tasks, web shopping, and generic skill use), so the gain reflects a broad performance uplift.

---

Source: "Skill Adapter: Personalized Modification and Online Routing of General‑Purpose Skills" (arXiv:2611.17007, 12 Nov 2026).

---

Source: Grafy Research Lab — Skill Adapter: Personalized Modification and Online Routing of General-Purpose Skills