Research map / Multi-agent systems and games / LLM agents
Efficient Skill Grounding via Code Refactoring with Small Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RECENT, a refactoring-centric framework where skills are stored as executable code with an ontology, and a small language model grounds them to new robots and environments via localized fill-in-the-middle edits instead of full regeneration.
Problem
Skill grounding for embodied agents across robot embodiments and dynamic environments normally requires regenerating whole skill code with LLMs, which is impractical on capacity-limited devices, while small language models lack the reasoning for reliable long-horizon control.
Approach
Separate invariant skill semantics from embodiment/environment-specific execution bindings in code; use an ontology to diagnose embodiment conflicts and an sLM with fill-in-the-middle to patch only the affected spans, plus in-situ unit-test-triggered patching of yet-to-be-executed code during execution.
Key results
- RECENT with Qwen2.5-Coder-7B reaches 73.00% SR (kinematic) and 82.50% SR (end-effector), versus 20.00% for CaP-CodeV-R1 on kinematic.
- RECENT achieves GO of 7.85% and 2.51%, far below full-regeneration baselines such as RepairAgent (54.17%) and CaP (102.01%) on kinematic.
- RECENT is within 6.58 pp SR of CaP-Codex (GPT-5.2-Codex) on average while reducing GO by 57.81 pp.
Surprising findings
- Full re-generation ablation drops SR to 26.67% and raises GO to 89.63%, showing localized editing matters more than the sLM alone.
Limitations
- Scoped to a shared, pre-built skill ontology; new capabilities require ontology extension.
- Ontology construction and skill validation rely on a cloud LLM and a fixed reference robot offline.
- Evaluated only in simulation with 5 robot embodiments and about 32 tasks, with real-world results in a separate section.
Experimental setup
Datasets: RLBench VIMA RoboGen
Benchmarks: Kinematic variation tasks (Panda to UR5/Sawyer, CoppeliaSim) End-effector variation tasks (Panda to Robotiq/vacuum, Genesis)
Metrics: SR GC GO EI IT
Baselines: CaP CaP-CodeV-R1 SCoT ProgPrompt RoboInspector RepairAgent Agentless CaP-Codex
Builds on: Code as Policies fill-in-the-middle skill ontology
Closest papers by meaning
- Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning (NeurIPS 2025 · similarity 0.91)
- CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation (ICML 2026 · similarity 0.89)
- Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks (ICML 2026 · similarity 0.89)
- SkillMaster: Toward Autonomous Skill Mastery in LLM Agents (NeurIPS 2026 · similarity 0.88)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.88)
- Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.