SignRound: Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs (arXiv:2309.05516)

SignRound paper by Cheng et al. EMNLP 2024 Findings. Introduces signed gradient descent (SignSGD) for optimizing weight rounding and clipping in 200 steps. Reports 6.91-33.22% absolute average accuracy improvement at 2 bits across 11 zero-shot tasks. Claims near-lossless 4-bit quantization in most scenarios. Published in EMNLP 2024 Findings.