Reproduce “SKIM: Pruning Large Language Model Agents via Selective Knowledge Informed Masking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Moonseok Choi, Jongwon Jeong, Minki Kang, Giung Nam, Juho Lee
The authors introduce SKIM to compress agents that produce API-call sequences under grammar constraints. SKIM uses tool-relevant context and choice points to guide structured pruning, then distills the compressed model on its own grammar-constrained responses. On Qwen3.5-2B, it reaches 64.5% structured full-match accuracy at 50% nominal sparsity, versus 56.2% for FLAP; this setting removes 34.14% of total parameters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.