Research map / Architectures and efficiency / Quantization

Zero-Shot Quantization via Weight-Space Arithmetic

NeurIPS 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes quantization vectors (QAT-minus-FT weight displacements) that can be added to a receiver model with a tuned scale to improve zero-shot 3-bit PTQ accuracy across vision and text classifiers, with proofs of well-definedness and a cosine-squared transfer law.

Problem

Post-training quantization at very low bit-widths (3-bit) severely degrades accuracy, while quantization-aware training is costly and must be repeated per task; the question is whether QAT-induced robustness can be reused across tasks without receiver-side training.

Approach

Define a quantization vector as the weight-space displacement between matched full-precision fine-tuned and QAT checkpoints on a donor task, add a scaled donor QV to a receiver checkpoint to improve its post-PTQ accuracy, and justify this with a gauge/well-definedness theorem and a local quadratic cosine-squared transfer law.

Key results

• Donor quantization vector patching improves 3-bit PTQ Top-1 accuracy by up to 60% on receiver tasks, without receiver-side QAT (per abstract).
• With validation-tuned scaling lambda, QV patching is reported to yield non-negative gains across donor-receiver pairs and closely approaches QAT+PTQ in vision, sometimes surpassing it.
• ImageNet is identified as a universal donor across receiver tasks.

Surprising findings

• Unit-scale (lambda=1) donor patching often causes destructive interference, which validation-tuned scaling largely removes.
• QV patching sometimes surpasses the receiver's own QAT+PTQ result.

Limitations

• Lambda selection requires a held-out validation split, so the method is not fully data-free.
• Evaluated only with a basic 3-bit weight-only per-channel PTQ setup; interaction with advanced PTQ techniques is untested.
• Theoretical cosine-squared law relies on a local quadratic model that may not hold globally.

Experimental setup

Datasets: Stanford Cars CIFAR-10 CIFAR-100 DTD EMNIST EuroSAT Fashion-MNIST FER2013 Flowers102 Food-101 GTSRB ImageNet

Benchmarks: Stanford Cars CIFAR-10 CIFAR-100 DTD EuroSAT GTSRB ImageNet MNIST SUN397 Emotion IMDB Banking77

Metrics: Top-1 accuracy transfer gain (Delta accuracy vs PTQ)

Baselines: vanilla 3-bit PTQ 3-bit QAT full-precision fine-tuning (FT)

Builds on: task arithmetic quantization-aware training straight-through estimator AdamW

Closest papers by meaning

• Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.90)
• SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models (NeurIPS 2026 · similarity 0.89)
• VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.89)
• AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.88)
• Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.88)
• SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning (ICLR 2025 · similarity 0.88)