Research map / Architectures and efficiency / Quantization
Post-training quantization of LLMs
Handles outliers and low-bit errors when quantizing large language models after training.
136 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Quantization, keyword cluster “post quantization · ptq · compression”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
25 papers in 2023–24 and 106 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 21.2% to 31.1%.
Open questions the data points to
- Papers report QuIP on WikiText2, Llama-2-70B, 2-bit (perplexity) anywhere from 3.91 to 5.9: comparisons against it depend on whose number you take.
- Papers report QTIP on WikiText2, Llama-2-7B, 2-bit (perplexity) anywhere from 5.91 to 6.82: comparisons against it depend on whose number you take.
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
- 2020 · Q-BERT: Hessian Based Ultra Low Precision Quantization of BERT (AAAI 2020 · 423 citations)
- 2021 · MQBench: Towards Reproducible and Deployable Model Quantization Benchmark (NeurIPS 2021 · 23 citations)
- 2022 · ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers (NeurIPS 2022 · 70 citations)
- 2023 · QLoRA: Efficient Finetuning of Quantized LLMs (NeurIPS 2023 · 824 citations)
- 2024 · OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models (AAAI 2024 · 72 citations)
- 2025 · ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · 16 citations)
- 2026 · HALO: Hardware-Aware Quantization with Low Critical-Path-Delay Weights for LLM Acceleration (AAAI 2026 · 1 citation)
Earliest papers here
- Q-BERT: Hessian Based Ultra Low Precision Quantization of BERT (AAAI 2020 · 423 citations)
- MQBench: Towards Reproducible and Deployable Model Quantization Benchmark (NeurIPS 2021 · 23 citations)
- ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers (NeurIPS 2022 · 70 citations)
What has been tried
59 papers here report 649 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 8%, hurt 19%, 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 |
|---|---|---|---|---|
| apply truncation without protecting weak columns | WikiText-2 · perplexity | 12.88 -> 48.26 | Hurt | OWQ: Outlier-Aware Weight Quantization for Efficient Fine-Tuning and Inference of Large Language Models (AAAI 2024 · 72 citations) |
| Add token pruning to AgileQ | Wikitext-2 · perplexity | 6.16 -> 6.09 | Helped | Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge (AAAI 2024 · 35 citations) |
| Use small weight groups with quantized statistics instead of larger groups with 16-bit statistics | Language modeling benchmarks · language modeling loss | Helped | SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · 26 citations) | |
| add salient-weight residual approximation | OPT-6.7B and LLaMA-7B · perplexity | Helped | BiLLM: Pushing the Limit of Post-Training Quantization for LLMs (ICML 2024 · 25 citations) | |
| add 0.45% sparse extraction of sensitive values and outliers | C4, 3-bit · perplexity | 7.75 -> 7.56 | Helped | SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · 23 citations) |
| Use block-size-256 fine-grained weight quantization instead of per-row quantization | OPT-6.7B average PPL · PPL | 12.10 -> 12.05 | Helped | Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · 22 citations) |
| quantize weights from 4-bit to 3-bit | LLaMA-30B, 1 GPU · latency | 20.7 -> 18.1 ms | Helped | LUT-GEMM: Quantized Matrix Multiplication based on LUTs for Efficient Inference in Large-Scale Generative Language Models (ICLR 2024 · 19 citations) |
| Apply bit balance strategy to W2A8 | WikiText2 · perplexity | 11.35 -> 7.59 | Helped | ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · 16 citations) |
| Add smoothing | WikiText2 · perplexity | 6.79 -> 6.28 | Helped | DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · 16 citations) |
| enable optimized quantization kernel | 18B, 64 GPUs · throughput | 19.73 -> 31.40 TFLOPs | Helped | ZeRO++: Extremely Efficient Collective Communication for Large Model Training (ICLR 2024 · 15 citations) |
| add LWC to LET | LLaMA-7B W4A4 · average perplexity | 16.97 -> 12.87 | Helped | OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models (ICLR 2024 · 14 citations) |
| Add channel disassembly to the quantized model | WikiText2 · perplexity | 1702.34 -> 19.34 | Helped | QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · 11 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 |
|---|---|---|---|---|
| QuIP | WikiText2, Llama-2-70B, 2-bit | perplexity | 3.91 – 5.9 | 2 papers, e.g. NeurIPS 2025 and ICML 2025 |
| QTIP | WikiText2, Llama-2-7B, 2-bit | perplexity | 5.91 – 6.82 | 2 papers, e.g. NeurIPS 2025 and ICML 2025 |
What reviewers will expect
Datasets and baselines named in the 53 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- WikiText2 · 25 of 53 papers
- C4 · 18 of 53 papers
- RedPajama · 9 of 53 papers
- GSM8K · 5 of 53 papers
- PIQA · 4 of 53 papers
- WinoGrande · 4 of 53 papers
- HellaSwag · 4 of 53 papers
- PTB · 3 of 53 papers
Baselines
- GPTQ · 20 of 53 papers
- OmniQuant · 14 of 53 papers
- RTN · 13 of 53 papers
- QuIP · 13 of 53 papers
- AWQ · 12 of 53 papers
- QuaRot · 12 of 53 papers
- SmoothQuant · 10 of 53 papers
- SpinQuant · 8 of 53 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 131 papers; from OpenAlex and the papers' affiliations).
- Huawei · 5 papers
- Tsinghua University · 5 papers
- Massachusetts Institute of Technology · 4 papers
- Seoul National University · 4 papers
- Microsoft · 3 papers
- ByteDance · 3 papers
- Alibaba Group · 3 papers
- The University of Hong Kong · 3 papers
- Google · 3 papers
- SqueezeBits Inc · 2 papers
Use it
Download BibTeX of 27 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
- Post-training weight quantization (100 papers)
- Post-training quantization of ViTs (105 papers)
- KV cache quantization (72 papers)
Elsewhere in the topic
- Mixed-precision quantization (92 papers)
- Low-precision training formats (58 papers)
- Ultra-low-bit LLM quantization (56 papers)
- Binary neural networks (27 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).