GLASS: Global-Local Aggregation for Inference-time Sparsification of LLMs

TL;DR

GLASS is a training-free framework that improves inference-time sparsification of LLMs by combining local prompt-specific and global model priors for more accurate neuron pruning, especially with short prompts.

Contribution

It introduces a novel global-local aggregation method for stable FFN pruning, enhancing performance in short-prompt, long-generation scenarios without additional training.

Findings

Achieves up to 45.10% lower perplexity

Reduces KL divergence by 25.73%

Provides significant on-device decoding speedup

Abstract

Inference-time sparsification is a promising path to deploy large language models (LLMs) on resource-constrained devices, yet existing training-free methods typically estimate feedforward network (FFN) neuron importance from the input prompt alone. We show this prompt-only signal is often unreliable, especially for short prompts and long-form decoding, leading to inaccurate masks and degraded generation fidelity. We propose GLASS, a plug-and-play, training-free framework that stabilizes dynamic FFN pruning by aggregating two complementary views of neuron criticality: local prompt-specific activations and a global model-intrinsic prior. GLASS fuses global and local signals via rank aggregation, yielding robust critical-neuron selection even when the prompt is short. We interpret GLASS as the maximum-a-posteriori consensus ranking under a permutation-based probabilistic model, providing a principled foundation for its weighted rank-aggregation rule. We apply GLASS to a diverse set of open-source LLMs, and show that it yields substantial improvements over prior training-free baselines in the challenging short-prompt, long-generation scenarios, achieving up to 45.10% lower perplexity and 25.73% lower KL divergence, while delivering significant on-device decoding speedup.