Research map / Architectures and efficiency / Quantization
Low-precision training formats
Designs 8-bit floating-point formats and accumulation schemes for low-precision neural network training.
58 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Quantization, keyword cluster “mixed precision · floating · point”
State History Benchmarks Tried Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
14 papers in 2023–24 and 12 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 11.9% to 3.5%.
No longer reported by papers here: ImageNet (top-1 accuracy, last in 2024); Cifar-10 (accuracy, last in 2024).
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
- 2017 · Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks (NeurIPS 2017 · 166 citations)
- 2018 · Mixed Precision Training (ICLR 2018 · 999 citations)
- 2019 · Hybrid 8-bit Floating Point (HFP8) Training and Inference for Deep Neural Networks (NeurIPS 2019 · 129 citations)
- 2020 · Rethinking Differentiable Search for Mixed-Precision Neural Networks (CVPR 2020 · 112 citations)
- 2021 · Any-Precision Deep Neural Networks (AAAI 2021 · 41 citations)
- 2022 · On-Device Training Under 256KB Memory (NeurIPS 2022 · 72 citations)
- 2023 · Training Transformers with 4-bit Integers (NeurIPS 2023 · 8 citations)
- 2024 · PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 2024 · 6 citations)
- 2025 · Analog Foundation Models (NeurIPS 2025 · 2 citations)
- 2026 · FlashOptim: Optimizers for Memory-Efficient Training (ICML 2026 · 0 citations)
Earliest papers here
- Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks (NeurIPS 2017 · 166 citations)
- ZipML: Training Linear Models with End-to-End Low Precision, and a Little Bit of Deep Learning (ICML 2017 · 107 citations)
- Mixed Precision Training (ICLR 2018 · 999 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)
7 papers, 2018–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 67.34 | Training Deep Neural Networks with 8-bit Floating Point Numbers (NeurIPS 2018 · 202 citations) |
| 2019 | 76.46 | Hybrid 8-bit Floating Point (HFP8) Training and Inference for Deep Neural Networks (NeurIPS 2019 · 129 citations) |
Cifar-10 · accuracy (higher is better)
4 papers, 2020–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 92.97 | FracTrain: Fractionally Squeezing Bit Savings Both Temporally and Spatially for Efficient DNN Training (NeurIPS 2020) |
| 2021 | 93.8 | Any-Precision Deep Neural Networks (AAAI 2021 · 41 citations) |
| 2022 | 94.8 | Is Integer Arithmetic Enough for Deep Learning Training? (NeurIPS 2022 · 7 citations) |
What has been tried
48 papers here report 536 measured changes (a component added, removed or swapped, compared with a baseline): helped 56%, no clear effect 17%, hurt 24%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use PoT-4 for pointwise and linear-8 for depthwise convolution weights instead of PoT for both | ImageNet · Top-1 accuracy | 64.50 -> 67.60 | Helped | PikeLPN: Mitigating Overlooked Inefficiencies of Low-Precision Neural Networks (CVPR 2024 · 6 citations) |
| Use pure softmax sampling for synthetic data generation rather than RGS | 1B-token training, benchmark average · accuracy | 65.66 -> 66.26 | Helped | Analog Foundation Models (NeurIPS 2025 · 2 citations) |
| Add a color coordinate to the bit-plane input | Kodak · iterations | 790 -> 1438 | Hurt | Towards Lossless Implicit Neural Representation via Bit Plane Decomposition (CVPR 2025 · 1 citation) |
| add BN | CIFAR10 · accuracy | 90.29 -> 92.37 | Helped | BOLD: Boolean Logic Deep Learning (NeurIPS 2024 · 1 citation) |
| Quantize P5, the input to the parameter-gradient computation | LLM pretraining · loss | up to 2% | Hurt | Scaling Laws for Floating–Point Quantization Training (ICML 2025 · 0 citations) |
| Add Dampen | ImageNet-1K validation, DeiT-T and DeiT-S · Top-1 accuracy | 59.75 -> 59.75; 71.03 -> 70.75 | Mixed | Oscillation-Reduced MXFP4 Training for Vision Transformers (ICML 2025 · 0 citations) |
| update matrix A | Alpaca, LLaMA-7B · time per step | 1.80 -> 1.94 | Hurt | FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · 0 citations) |
| Use dual-stage rather than 1-stage fine-tuning | ImageNet · Top-1 accuracy | 74.15% -> 76.22% | Helped | Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · 0 citations) |
| Add MCF for optimizer states in addition to parameter MCF | GPT-125M, β₂=0.999 · validation perplexity | 15.80 -> 15.18 (Global BatchSize=1024) | Helped | Collage: Light-Weight Low-Precision Strategy for LLM Training (ICML 2024 · 0 citations) |
| Increase β from 5 to 15 for linear-layer RTN | Zero-shot ARC-c · accuracy | 39.3 -> 43.0 | Helped | IM-Unpack: Training and Inference with Arbitrarily Low Precision Integers (ICML 2024 · 0 citations) |
| Add Dynamic Range Expansion to optimizer-state quantization | Optimizer quantization · MSE | +1.63× MSE reduction | Helped | COAT: Compressing Optimizer states and Activations for Memory-Efficient FP8 Training (ICLR 2025) |
| use AMPA instead of full INT6 training | CIFAR-10 · accuracy degradation (%) | Helped | AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024) |
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 26 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 3 papers
- Seoul National University · 2 papers
- Rice University · 2 papers
- The Chinese University of Hong Kong · 2 papers
Use it
Download BibTeX of 39 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
- Mixed-precision quantization (92 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).