Pruning Smarts: Dense LLMs Cliff, MoE Resilience in Smart Homes
A study examines how pruning affects smart-home LLMs across architectures, revealing mixed gains and losses with limits.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Congjing Zhang, Vashishtha Patil, Henning Lange, Usman Aleem
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
The study pruned four LLMs (dense, dense-hybrid, MoE) and evaluated over 19,500 smart-home instances. Dense models showed sharp degradation after modest pruning; MoE models tolerated more pruning, while grounding and value prediction degraded before higher-level intent.
Why it matters
Pruning can lower costs, but care is needed to preserve device grounding and precise value prediction in smart homes.
Dense LLMs lose grounding precision early, especially for device and value predictions, even when overall accuracy looks okay. MoE models resist pruning better, but all architectures can shift toward refusing to act with heavy pruning.
The key takeaway is to test not just overall accuracy but how well a model handles real, grounded tasks in a changing home context, especially for partially executable or multi-device requests.
Original sources · 1
- What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity ↗arXiv · 2026-09-15
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
Limited to Qwen family models; not all pruning methods apply equally across architectures; results may not generalize beyond smart-home tool calling.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.