Efficient Skill Grounding via Code Refactoring with Small Language Models

A new framework called RECENT allows small language models to ground reusable robot skills by refactoring code instead of rewriting it, according to a paper submitted on 6 Jun 2026 [1]. The approach matches the task performance of systems that rely on large language models while running on smaller, more practical models [1][2]. Embodied agents — robots operating in physical, dynamic environments — depend on reusable skill libraries to perform long sequences of tasks. Even small differences in a robot's body or surroundings can break a skill, a problem known as skill grounding [1][2]. Large language models (LLMs) have been used to handle this adaptation, but their size makes them impractical for many real-world deployments where compute, latency, or connectivity are constrained [1]. Small language models (sLMs) would be a natural alternative, yet they have lacked the capacity to reliably ground skills for extended operations [1][2]. The RECENT framework, short for a refactoring-centric agent, addresses this gap by separating a skill's semantic intent from the low-level commands that bind it to a specific robot or environment [1][2]. Skills are represented as executable code. When a skill needs to be adapted, RECENT does not regenerate the program from scratch. It performs localized refactoring, altering only the execution bindings while preserving the control structure that encodes what the skill is meant to do [1][2]. In evaluations across multiple robot embodiments and dynamic environments, RECENT achieved the highest performance among sLM-based Code-as-Policies methods and equaled the task success rate of LLM-based CaP systems [1][2]. The paper notes that this result holds across all tested scenarios, pointing to robust long-horizon performance when an sLM is paired with the framework [1][2]. The work was posted to arXiv on 6 Jun 2026 under the title "Efficient Skill Grounding via Code Refactoring with Small Language Models" [1]. The authors frame the contribution as a step toward making capable embodied agents viable without dependence on large, cloud-hosted models [1][2].