Parametric Skills

A new framework called ParametricSkills aims to convert free-form textual skill instructions into model parameters at test time, bypassing the need for large language models to read and follow lengthy prompts, according to a paper submitted in 2026 [1]. Large language models, or LLMs, are machine learning models with many parameters trained on vast amounts of text for tasks such as language generation [8]. Their ability to solve problems often depends on skills, which the paper describes as textual recipes encoding mature problem-solving experience [1]. A persistent limitation is that a model must comprehend and adhere to these skill instructions within its context window, a challenge that grows in complex, long-context scenarios [1]. To address this, the authors propose ParametricSkills, a framework that encodes both the content of a skill and the method for exploiting it into low-rank adaptation, or LoRA, adapters [1]. The system first constructs a large-scale skill library and synthesizes single-turn and multi-turn skill exploitation trajectories using a tool called OpenCode [1]. A hypernetwork is then trained to receive textual skills and convert them into these adapters, enabling what the paper calls context-free skill exploitation at test time [1]. In experiments across six complex software engineering subtasks, ParametricSkills outperformed standard in-context learning by an average of 6.44 points as judged by DeepSeek-V4-Flash [1]. The framework also achieved higher BERT Score and F1 score metrics, confirming its effectiveness on the tested benchmarks [1]. The paper further notes that because parametric skills are inherently accumulative, they offer a preliminary avenue toward test-time continual learning, where a model could build on previously acquired skills without reprocessing their full textual descriptions [1]. The concept of converting descriptive inputs into operational parameters has parallels in other domains. In computer-aided design, parametric modeling introduced by PTC Inc. in 1988 allowed engineers to define solid models through associative, feature-based parameters rather than static geometry [2][3]. The underlying principles of solid modeling emphasize physical fidelity and support the creation, visualization, and annotation of digital models of physical objects [3]. While the ParametricSkills framework operates on language rather than three-dimensional shapes, both approaches share the goal of making complex instructions directly executable by a system without requiring it to reinterpret the original specification each time. The paper was submitted on 29 June 2026 to the arXiv preprint server in the Computation and Language category [1].