First-Order Error Matters: Accurate Compensation for Quantized Large Language Models

Abstract: Post-training quantization (PTQ) offers an efficient approach to compressing LLMs, significantly reducing memory access and computational costs. Existing compensation-based weight calibration methods often rely on a second-order Taylor expansion to model quantization error, under the assumption that the first-order term is negligible in well-trained full-precision models. However, we reveal that the progressive compensation process introduces accumulated first-order deviations between latent weights and their full-precision counterparts, making this assumption fundamentally flawed. To address this, we propose FOEM, a novel PTQ method that explicitly incorporates first-order gradient terms to improve quantization error compensation. FOEM approximates gradients by directly computing the difference between latent and full-precision weights, avoiding the high cost and limited generalization of backpropagation-based gradient computation. This approach introduces minimal additional computational overhead. Moreover, FOEM leverages precomputed Cholesky factors to efficiently recover the inverse of Hessian submatrices in real time. Extensive experiments across a wide range of models and benchmarks demonstrate that FOEM consistently outperforms the classical GPTQ method. In 3-bit weight-only quantization, FOEM reduces the perplexity of Llama3-8B by 89.6%, and improves the 5-shot MMLU accuracy of Llama3-70B from 51.7% to 74.9%, approaching the full-precision performance of 78.6%. Furthermore, FOEM can be seamlessly integrated with advanced techniques such as GPTAQ and SpinQuant, yielding additional improvements under the challenging W4A4KV4 setting, and further narrowing the accuracy gap with full-precision baselines beyond what current state-of-the-art methods achieve. The code is available at https://github.com/Xingyu-Zheng/FOEM.