Research map / Architectures and efficiency / Quantization
Mixed-precision quantization
Assigns bit-widths per layer for low-bit network training and inference with learnable quantizers.
92 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Quantization, keyword cluster “mixed precision · floating · point”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
9 papers in 2023–24 and 13 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 7.6% to 3.8%.
Open questions the data points to
- Papers report DOREFA on ImageNet (Top-1 accuracy) anywhere from 47.7 to 69.9: comparisons against it depend on whose number you take.
No longer reported by papers here: ImageNet (Top-1 accuracy, last in 2023); CIFAR-10 (top-1 accuracy (%), last in 2022); ImageNet (Top-5 accuracy, last in 2021).
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Fixed Point Quantization of Deep Convolutional Networks (ICML 2016 · 599 citations)
- 2017 · Trained Ternary Quantization (ICLR 2017 · 836 citations)
- 2018 · Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference (CVPR 2018 · 3,940 citations)
- 2019 · HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · 963 citations)
- 2020 · ZeroQ: A Novel Zero Shot Quantization Framework (CVPR 2020 · 386 citations)
- 2021 · Network Quantization With Element-Wise Gradient Scaling (CVPR 2021 · 102 citations)
- 2022 · Learnable Lookup Table for Neural Network Quantization (CVPR 2022 · 53 citations)
- 2023 · OMPQ: Orthogonal Mixed Precision Quantization (AAAI 2023 · 26 citations)
- 2024 · Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · 17 citations)
- 2025 · JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · 6 citations)
- 2026 · InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · 1 citation)
Earliest papers here
- Fixed Point Quantization of Deep Convolutional Networks (ICML 2016 · 599 citations)
- Trained Ternary Quantization (ICLR 2017 · 836 citations)
- Deep Learning With Low Precision by Half-Wave Gaussian Quantization (CVPR 2017 · 484 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ImageNet · Top-1 accuracy (higher is better)
12 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 72.23 | Deep Neural Network Compression With Single and Multiple Level Quantization (AAAI 2018 · 112 citations) |
| 2020 | 78.2 | Fixed-Point Back-Propagation Training (CVPR 2020 · 44 citations) |
ImageNet, ResNet18 · Top-1 accuracy (higher is better)
9 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 70.9 | Additive Powers-of-Two Quantization: An Efficient Non-uniform Discretization for Neural Networks (ICLR 2020 · 118 citations) |
| 2021 | 71.56 | HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · 38 citations) |
ImageNet, ResNet-50 · Top-1 accuracy (higher is better)
6 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 76.6 | Additive Powers-of-Two Quantization: An Efficient Non-uniform Discretization for Neural Networks (ICLR 2020 · 118 citations) |
| 2021 | 77.58 | HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · 38 citations) |
CIFAR-10 · accuracy (higher is better)
5 papers, 2018–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 91.75 | Deep Neural Network Compression With Single and Multiple Level Quantization (AAAI 2018 · 112 citations) |
| 2020 | 95 | Towards Unified INT8 Training for Convolutional Neural Network (CVPR 2020 · 162 citations) |
CIFAR-10 · top-1 accuracy (%) (higher is better)
5 papers, 2019–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 94.76 | SeerNet: Predicting Convolutional Neural Network Feature-Map Sparsity Through Low-Bit Quantization (CVPR 2019 · 82 citations) |
ImageNet · Top-5 accuracy (higher is better)
5 papers, 2018–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 91 | Deep Neural Network Compression With Single and Multiple Level Quantization (AAAI 2018 · 112 citations) |
| 2021 | 92.7 | Optimizing Information Theory Based Bitwise Bottlenecks for Efficient Mixed-Precision Activation Quantization (AAAI 2021 · 2 citations) |
What has been tried
79 papers here report 825 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 8%, hurt 22%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add interval analysis to ILP | MNIST test, FC1-100, r=4 · total execution time (s) | 332.48 -> 93.52 | Helped | Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · 17 citations) |
| remove the 2-bit option from weight-sharing quantization | ImageNet, 6-bit weight-sharing model · Top-1 accuracy | 69.2 -> 70.4 | Helped | Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · 7 citations) |
| select the most important activation channels for sparse quantization rather than using heterogeneous sampling | CIFAR-100 · Top-1 accuracy | 55.704 -> 65.863 | Helped | JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · 6 citations) |
| bit-meta training (meta-state weights) before bit search | MobileNet-v2 activation distribution · activation consistency across bit-widths | Helped | MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · 4 citations) | |
| Add bias correction | Activation alignment, averaged across bit-widths · MAE | 0.755 to 0.667 | Helped | Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 1 citation) |
| Add ASGA to EdMIPS | CIFAR10 proxy · Top-1 accuracy | 66.4 -> 67.9 | Helped | Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · 0 citations) |
| Remove int8 and int4 loss terms, optimizing only int2 (S.P. Matryoshka Quantization) | int2 FFN · Task Avg. | 52.37 -> 53.42; 63.35 -> 64.02; 62.75 -> 63.58 | Helped | Matryoshka Quantization (ICML 2025 · 0 citations) |
| Add improved training recipe | BigANN1M, 8 bytes · training time | 127:38 -> 14:53 | Helped | Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · 0 citations) |
| Add EP-init to A2Q at P=12 | ImageNet-1K · top-1 accuracy | 35.5% -> 42.7% | Helped | A2Q+: Improving Accumulator-Aware Weight Quantization (ICML 2024 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| DOREFA | ImageNet | Top-1 accuracy | 47.7 – 69.9 | 2 papers, e.g. CVPR 2018 and CVPR 2022 |
| LSQ | ImageNet | Top-1 accuracy (%) | 71.1 – 76.7 | 2 papers, e.g. ICLR 2022 and CVPR 2022 |
What reviewers will expect
Datasets and baselines named in the 7 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 3 of 7 papers
- CIFAR-100 · 2 of 7 papers
- ImageNet · 2 of 7 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 22 papers; from OpenAlex and the papers' affiliations).
- CATL · 2 papers
- Google · 2 papers
- Tsinghua University · 2 papers
- Northeastern University · 2 papers
Use it
Download BibTeX of 37 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Low-precision training formats (58 papers)
- Ultra-low-bit LLM quantization (56 papers)
- Binary neural networks (27 papers)
Elsewhere in the topic
- Post-training quantization of LLMs (136 papers)
- Post-training quantization of ViTs (105 papers)
- Post-training weight quantization (100 papers)
- KV cache quantization (72 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).