Technical Summary: ParametricSkills

Problem Statement

Agentic Large Language Models (LLMs) rely heavily on skills—structured textual recipes encoding problem-solving experience—to handle complex real-world tasks like software engineering (SWE) and scientific research. However, the prevailing paradigm of representing these skills as in-context text faces two fundamental limitations:

1. Instruction Following Bottleneck: Models struggle to comprehend and follow complex, long-context textual instructions. Vital instructions are often lost or difficult to locate, leading to suboptimal task outcomes, particularly as model size decreases.
2. Decoupling of Evolution and Learning: Optimizing or evolving skills in text space (e.g., rewriting natural language documents) is decoupled from the model's intrinsic learning. While text-based methods can improve downstream performance, they do not update the model's parameters, leaving the model's core capabilities unoptimized.

Methodology: ParametricSkills

The authors propose ParametricSkills, a framework that converts free-form textual skills into parametric adapters (specifically LoRA adapters) at test time using a hypernetwork. This approach eliminates the need for context-window consumption and couples skill evolution directly with model learning.

Core Architecture

The system utilizes a hypernetwork ($h_\theta$) that takes a textual skill description ($s$), encodes it via a frozen text encoder ($g(\cdot)$) into a representation ($z$), and generates LoRA parameters ($\phi$) in a single forward pass:
$$\phi = h_\theta(g(s))$$
These parameters are merged into the base model ($W_0 + \Delta W$), creating a skill-conditioned model without modifying the base weights or consuming context tokens.

Training Pipeline

The training process consists of three stages:

1. Skill-Reconstruction Pretraining: The hypernetwork is trained on a large skill library using three self-supervised objectives to ensure faithful encoding of skill content:
   • Full Reconstruction: Reconstructing the entire context from the full skill.
   • Prefix Completion: Inferring the full context from a truncated prefix.
   • Section-Level Completion: A cloze-style task where a functional section is removed, requiring bidirectional reasoning to reconstruct the missing component.
2. Skill-Exploitation Fine-tuning: The hypernetwork is fine-tuned on synthesized agentic trajectories (single-turn and multi-turn). This shifts the focus from merely compressing content to learning the methodology of skill exploitation, grounding the parameters in concrete behavioral steps (planning, tool use, verification).
3. Self-Evolving Loop: Skills can be refined in text space (via a generator-verifier loop) and recompiled into parameter space. For continual learning, successful task trajectories are summarized, converted to parameters, and merged into a global parametric skill via Exponential Moving Average (EMA).

Data Curation

The authors constructed a large-scale dataset comprising:

• Skill Library: 45.8k high-quality skills from web repositories (e.g., GitHub, npm) and mined from successful agent trajectories across 13 domains (e.g., Software, AI/LLM, Security).
• Trajectories: Synthetic single-turn (QA-style and realistic scenarios) and multi-turn (sandbox-based) exploitation trajectories generated using the OpenCode sandbox.

Skill Merging

To handle complex tasks requiring multiple skills, the framework supports merging multiple LoRA adapters. The authors propose a rank-concatenation method where weighted updates of individual adapters are concatenated, outperforming simple factor-wise linear combination.

Key Contributions

1. Large-Scale Skill Infrastructure: Construction of a 45.8k skill library and a corresponding dataset of agentic trajectories with rigorous quality control.
2. Parametric Skill Framework: Introduction of a test-time parameterization framework that internalizes skill exploitation methodologies into context-free parameters, addressing the long-context instruction-following bottleneck.
3. Empirical Superiority: Comprehensive evaluation demonstrating that ParametricSkills outperforms in-context learning baselines and text-to-LoRA baselines (SHINE) on complex SWE tasks.
4. Continual Learning Pathway: Demonstration that parametric skills can be seamlessly integrated with self-evolving pipelines and accumulated via EMA, offering a viable avenue for test-time continual learning without backpropagation.

Experimental Results

Evaluated on six complex SWE subtasks (e.g., Feature/Patch, Robustness Fix, Refactoring) using DeepSeek-V4-Flash as the judge:

• Performance: ParametricSkills achieved an average LLM judge score of 64.09, outperforming the In-Context baseline (57.65) by 6.44 points and the No-Skills baseline (50.21) by a significant margin.
• Metrics: It also achieved higher BERT Scores (+1.17) and F1 Scores (+5.53%) compared to the In-Context baseline.
• Baseline Comparison: The text-to-LoRA baseline (SHINE) failed to surpass the In-Context baseline (56.36 vs 68.12), highlighting that naive text-to-parameter conversion is insufficient without specific skill-exploitation training.
• Continual Learning: In online accumulation experiments on HumanEval, the "Online continual merge" setting achieved a 51.61% pass rate, significantly outperforming the independent self-evolution baseline (38.71%).

Significance

The paper posits that representing skills in parameter space rather than text space fundamentally changes the agentic paradigm. By converting skills into LoRA adapters, the framework:

• Removes the dependency on the model's ability to locate and follow intricate instructions in long contexts.
• Makes skill evolution inherently coupled with model learning, as improving the skill directly improves the model's parameters.
• Provides a preliminary yet promising avenue for test-time continual learning, allowing models to accumulate experience from multiple trajectories into a global parametric prior without the computational overhead of traditional fine-tuning.