Research map / Architectures and efficiency / Quantization
FlexRound: Learnable Rounding based on Element-wise Division for Post-Training Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace AdaRound with division-based FlexRound | ImageNet weight-only PTQ · ImageNet, 4-bit weights, MobileNetV2 | Top-1 accuracy: 69.46 -> 70.82 | Helped |
| replace AdaRound with division-based FlexRound | ImageNet weight-only PTQ · ImageNet, 2-bit weights, ResNet-50 | Top-1 accuracy: 48.47 -> 63.67 | Helped |
| replace AdaRound with FlexRound | QDrop PTQ · ImageNet, 4-bit weights and activations, MobileNetV2 | Top-1 accuracy: 65.42 -> 66.66 | Helped |
| replace AdaRound with FlexRound | QDrop PTQ · GLUE MNLI, BERT_BASE | matched accuracy: 83.69 -> 84.53 | Helped |
| replace AdaRound with FlexRound | QDrop PTQ · GLUE QNLI, BERT_LARGE | accuracy: 92.24 -> 92.04 | Hurt |
| replace AdaRound with FlexRound | QDrop PTQ · WikiText2, GPT-Neo 125M | perplexity: 30.52 -> 24.30 | Helped |
| replace AdaRound with FlexRound | QDrop PTQ · WebNLG, GPT-2 M, All | BLEU: 54.05 -> 55.06 | Helped |
| replace AdaRound with FlexRound | QDrop PTQ · LLaMA-33B, BoolQ | accuracy (%): 64.86 -> 69.08 | Helped |
| learn the quantization grid size jointly with rounding rather than fixing it (ablation) | FlexRound · ImageNet, weights-only 4-bit, ResNet-18 | Top-1 accuracy: 70.09 -> 70.28 | Helped |
| add channel tensors s3 and s4 (ablation) | FlexRound · ImageNet, weights-only 4-bit, MobileNetV2 | Top-1 accuracy: 70.51 -> 70.82 | Helped |
| combine AdaQuant with FlexRound instead of using FlexRound alone (ablation) | ImageNet weight-only PTQ · ImageNet, 4-bit weights, MobileNetV2 | Top-1 accuracy: 70.82 -> 70.15 | Hurt |
| replace AdaRound with FlexRound | QDrop PTQ · GLUE SST-2, GPT-Neo 125M | accuracy: 84.75 -> 83.03 | Hurt |
Closest papers by meaning
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.88)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.88)
- FlatQuant: Flatness Matters for LLM Quantization (ICML 2025 · similarity 0.88)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.88)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.87)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.