Modern deep learning models—Large Language Models (LLMs), Vision Transformers (ViTs), and Diffusion Models—are massive. Deploying them on edge devices requires quantization: converting 32-bit floating-point weights into low-bit integers (like 4-bit or even 2-bit). However, as precision drops, model accuracy typically collapses.
The research direction emerging from these papers focuses on Post-Training Quantization (PTQ) techniques that bridge this accuracy gap without the massive computational cost of full retraining. By intelligently managing outliers and optimizing weight distributions during a brief "calibration" phase, researchers are making ultra-low-bit models viable for real-world hardware.
The Outlier Problem
The primary enemy of low-bit quantization is the "outlier"—a small number of weights or activations with values significantly larger than the rest. In a uniform quantization scheme, these outliers force the quantization grid to stretch wide, leaving the majority of "normal" values to be lumped into just one or two bins, destroying the model's representational power.
To combat this, researchers have moved beyond simple rounding. Zechun Liu and Igor Fedorov in SpinQuant propose using learned rotation matrices to "smear" these outliers across multiple dimensions, making the distribution more uniform and easier to quantize. Similarly, Song Han and Zhijian Liu introduced ParoQuant, which uses pairwise rotations to eliminate outliers specifically for reasoning-heavy LLMs.
Error Compensation and Reconstruction
If we cannot eliminate all quantization errors, we can at least try to ensure they don't accumulate. Traditional PTQ treats each layer in isolation, but newer methods look at the model holistically.
Ian Colbert and colleagues developed Qronos, which explicitly corrects errors by looking at how quantization in one layer affects the inputs to the next. For Vision Transformers, Zhuguanyu Wu and Jiayi Zhang proposed APHQ-ViT, which uses Hessian-based reconstruction—a mathematical way of identifying which weights are most "sensitive" to change—to guide the quantization process.
Strategic Mixed Precision
Not all weights are equally important. A recurring strategy is to keep a tiny fraction of "high-impact" weights in high precision while aggressively compressing the rest.
Yan Gu and Yu Shao demonstrated in GWQ that preserving just the top 1% of weights (identified via gradients) in FP16 allows the remaining 99% to be pushed to 3 bits with minimal loss. QEFT by Changhun Lee and Younghyun Cho takes this further by reordering weight columns to create a hardware-friendly mixed-precision format that speeds up both inference and fine-tuning.
The Calibration Trade-off
While these methods avoid the weeks of compute required for Quantization-Aware Training (QAT)—as described in foundational works like Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference—they are not "free."
A central theme across this research is the calibration overhead. Methods like MPPQ and BTC-LLM require a one-time optimization phase that can take several GPU hours. This phase involves running a small set of "calibration data" through the model to find optimal scaling factors, rotation angles, or codebooks. As Jiaqi Zhao notes in PTQ1.61, this preprocessing is the price we pay for pushing the limits of sub-2-bit compression.
Hardware-Algorithm Co-design
Finally, the effectiveness of quantization is ultimately measured by its performance on silicon. Renjie Wei and Meng Li argue in FPQVAR that we should move away from standard integer quantization (INT) toward low-bit floating-point (FP) formats, which better match the natural distribution of neural network weights. By co-designing the algorithm with specialized FPGA hardware, they achieved 4-bit generation that is 3.1× faster than previous integer-based systems.