SKIM cuts agent skills to 30-60% of their length for a 1-2 point accuracy hit

Notes on Adaptive Multi-Resolution Procedural Knowledge Compression for Large Language Models (arXiv:2606.12203) — Changyue Wang, Weihang Su, Qingyao Ai, Yichen Tang, Runzhong Qiao, Xuancheng Li, Min Zhang, Yiqun Liu · June 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SKIM (SKIll coMpression) replaces a reusable agent skill's full instructions with a small set of learned soft tokens, so the skill no longer has to be pasted into every prompt in full. A compressor model reads the skill text alongside a fixed budget of learnable slot tokens, an MLP projector maps the result into the target model's embedding space, and a LoRA adapter teaches the target LLM to treat those soft tokens as if it had read the source text directly. Training moves from general to specific across three stages: it first reconstructs 60.9k collected skills to build general procedural knowledge representations, warms up next on procedural question-answering over 214.3k WikiHow-derived QA examples, and finally aligns on roughly 60k skill-conditioned QA examples via LoRA fine-tuning with the target model. Skill complexity varies enormously, so before deployment SKIM runs what the paper calls offline resolution selection: generate diagnostic questions from a skill, check how well answers at each token budget (256, 512, or full text) match the full-text answer, and lock in the smallest budget that clears a fidelity bar — all computed once, offline, before the skill is ever deployed.

Across five benchmark suites — BigCodeBench (code), CHAMP (math), LogicBench (logic), TheoremQA (theorem proving), ToolQA (tool use) — and two target models, Qwen3-8B and Phi-4, SKIM holds onto most of the task accuracy while the paper reports compressing skills to 30-60% of their original length overall. Looking at total per-task context in the adaptive setting: BigCodeBench shrinks from 4,676 to 3,172 tokens for a 2.4-point drop, 49.30% to 46.93%; LogicBench shrinks from 1,089 tokens to 438 for only a 1.6-point drop. General-purpose compressors don't come close: on the single-skill ToolQA benchmark, where SKIM, ICAE, and 500xCompressor all compress to the same 256 tokens, SKIM holds 35.66% accuracy against ICAE's 7.13% and 500xCompressor's 0.00%. On BigCodeBench, ICAE compressed to a flat 256 tokens manages 29.47% and 500xCompressor collapses to 0.18%, while SKIM's 256-token-per-skill setting — 708 tokens total, since BigCodeBench instances average multiple skills — reaches 44.21%.