Search papers, labs, and topics across Lattice.
This paper investigates token-level redundancy in Large Speech Language Models (LSLMs) by using layer-wise oracle interventions to reveal that deep layers exhibit extreme redundancy compared to shallow layers. Based on this finding, they introduce Affinity Pooling, a training-free, similarity-based token merging mechanism applied at both input and deep layers to compress speech representations. Experiments on three tasks show that Affinity Pooling reduces prefilling FLOPs by 27.48\% while maintaining competitive accuracy, leading to significant efficiency gains in memory and time-to-first-token.
LSLMs can be significantly compressed without sacrificing accuracy by aggressively merging redundant tokens in deeper layers, challenging the need for fully distinct token representations.
Large Speech Language Models (LSLMs) typically operate at high token rates (tokens/s) to ensure acoustic fidelity, yet this results in sequence lengths that far exceed the underlying semantic content, incurring prohibitive inference costs. In this paper, we empirically revisit the necessity of such granular token-level processing. Through layer-wise oracle interventions, we unveil a structured redundancy hierarchy: while shallow layers encode essential acoustic details, deep layers exhibit extreme redundancy, allowing for aggressive compression. Motivated by these findings, we introduce Affinity Pooling, a training-free, similarity-based token merging mechanism. By strategically applying this method at both input and deep layers, we effectively compress speech representations without compromising semantic information. Extensive evaluations across three tasks demonstrate that our approach reduces prefilling FLOPs by 27.48\% while maintaining competitive accuracy. Practical deployment further confirms significant efficiency gains, yielding up to $\sim$1.7$\times$ memory savings and $\sim$1.1$\times$ faster time-to-first-token on long utterances. Our results challenge the necessity of fully distinct token representations, providing new perspectives on LSLM efficiency.
