Fitting Is Not Enough: Smoothness in Extremely Quantized LLMs
For researchers and practitioners deploying LLMs with extreme quantization, this work highlights smoothness preservation as a critical design consideration that can improve generation quality beyond traditional numerical accuracy metrics.
This paper identifies that extremely quantized LLMs suffer from systematic smoothness degradation beyond numerical precision loss, which leads to sparser decoding trees and degraded generation quality. They show that preserving smoothness brings additional gains beyond numerical accuracy in both post-training quantization and quantization-aware training.
Large language models (LLMs) achieve strong performance but incur high deployment costs, motivating extremely low-bit but lossy quantization. Existing quantization algorithms mainly focus on improving the numerical accuracy of forward computation to eliminate performance degradation. In this paper, we show that extremely quantized LLMs suffer from systematic smoothness degradation beyond numerical precision loss. Through a smoothness proxy, we observe that such degradation becomes increasingly severe as the quantization bit-width decreases. Furthermore, based on sequence neighborhood modeling, we find that quantized models exhibit a rapid reduction of effective token candidates within the prediction neighborhood, which directly leads to a sparser decoding tree and degraded generation quality. To validate it, we introduce a simple smoothness-preserving principle in both post-training quantization and quantization-aware training, and demonstrate that preserving smoothness brings additional gains beyond numerical accuracy. The core goal of this paper is to highlight smoothness preservation as an important design consideration for future extreme quantization methods. Code is available at https://github.com/xuyuzhuang11/FINE.