SKIM: An Adaptive Multi-Resolution Procedural Knowledge Compression Framework

This article introduces SKIM, an adaptive multi-resolution soft token compression framework for LLM procedural skills, which can compress skill text to 30%-60% of its original length while maintaining task performance superior to existing compression methods. SKIM is specifically designed for procedural knowledge, addressing the context inflation problem of LLMs and improving reasoning efficiency. Original author: bebr2, source: arXiv, release date: 2026-06-10, open-source code available on GitHub: https://github.com/bebr2/SKIM.

Large language models (LLMs) are evolving into intelligent agents, requiring loading multiple skills which leads to context inflation, increasing pre-filling costs and inference latency. Existing compression methods target factual knowledge and fail to preserve structural information such as logical dependencies, tool protocols, and conditional branches of procedural knowledge, easily breaking key dependencies required for skill execution.

SKIM proposes three core requirements for effective skill compression: 1. Preserve logical dependencies: Ensure that the logical relationships of workflows and tool protocols are maintained after compression; 2. Support lightweight offline compression: Adapt to rapid iteration of community skills without expensive retraining; 3. Adapt to different complexities: Adjust compression rates adaptively based on skill complexity (steps, nesting, branches, etc.).

SKIM is an adaptive multi-resolution soft token compression framework: 1. Soft token mechanism: Convert text into continuous vector representations with high information density, differentiable optimization, and preserved semantic structure; 2. Adaptive multi-resolution strategy: Select compression resolution through complexity evaluation and dynamically generate different numbers of soft tokens; 3. Offline process: Skill parsing → dependency graph construction → soft token generation → quality verification.

SKIM achieves a compression rate of 30%-60% (depending on skill complexity), with task performance superior to uncompressed original skills and existing methods. Advantages include: better preservation of procedural knowledge, higher compression efficiency, and lower computational overhead. Inference efficiency is significantly improved: reduced pre-filling time, lower memory usage, and improved end-to-end latency.

SKIM is suitable for: 1. Intelligent agent platforms (e.g., GPTs, Claude Artifacts): Reduce skill loading overhead and support simultaneous loading of multiple skills; 2. Enterprise knowledge bases: Efficiently integrate standard operating procedures, troubleshooting guides, etc.; 3. Community skill ecosystems: Lightweight offline compression adapts to rapidly iterating open-source skill libraries.

Current limitations: Domain adaptability (needs tuning for vertical fields like healthcare/legal), interpretability (soft tokens are less easy to debug than natural language), cross-model compatibility (bound to specific architectures). Future directions: Multi-modal skill compression, runtime dynamic adaptive compression rates, federated compression to protect privacy.

SKIM code has been open-sourced (GitHub link: https://github.com/bebr2/SKIM), providing a complete framework, pre-trained checkpoints, benchmark datasets, and documentation. SKIM is an important advancement in the field of procedural knowledge compression, providing key infrastructure support for large-scale LLM skill ecosystems.