Research map / Architectures and efficiency / Quantization
QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models
Jingxuan Zhang, Yunta Hsieh, Zhongwei Wan, Haokun Lin, Xin Wang, Ziqi Wang, Yingtie Lei, Mi Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| quantize DiT attention projections too (all DiT linear layers) instead of only the MLP (ablation) | QuantVLA layout on pi0.5 · LIBERO average | success rate: 95.4% -> 76.3% | Hurt |
| quantize DiT attention projections too instead of only the MLP (ablation) | QuantVLA layout on GR00T N1.5 · LIBERO average | success rate: 82.5% -> 70.0% | Hurt |
| quantize only the DiT (all layers) instead of LLM only (ablation) | W4A8 without ATM/OHB on pi0.5 · LIBERO average | success rate: 96.5% -> 71.6% | Hurt |
| add ATM and OHB calibration to the selective layout (ablation) | QuantVLA on pi0.5 · LIBERO average | success rate: 95.4% -> 97.6% | Improved |
| add ATM and OHB calibration to the selective layout (ablation) | QuantVLA on GR00T N1.5 · LIBERO average | success rate: 82.5% -> 88.0% | Improved |
| use QuantVLA instead of DuQuant on LLM+DiT | pi0.5 W4A8 · LIBERO average | success rate: 76.3% -> 97.6% | Improved |
| use QuantVLA instead of DuQuant on LLM+DiT | GR00T N1.5 W4A8 · LIBERO average | success rate: 70.0% -> 88.0% | Improved |
| quantize DiT MLP in addition to LLM (full QuantVLA vs LLM only) | QuantVLA on pi0.5 · Memory (GB) | success rate: 1.58 -> 1.28 | Improved |
| quantize DiT MLP in addition to LLM (full QuantVLA vs LLM only) | QuantVLA on GR00T N1.5 · LIBERO average | success rate: 87.0% -> 88.0% | Improved |
| lower activation precision from W4A8 to W4A4 | QuantVLA on pi0.5 · LIBERO average | success rate: 97.6% -> 95.3% | Mixed |
| increase denoising steps from 8 to 16 | QuantVLA on GR00T N1.5 · LIBERO average | success rate: 88.0% -> 88.5% | No clear effect |
| use QuantVLA instead of SmoothQuant | GR00T W4A8 · Pick-and-Can | successes out of 50: 16 / 50 -> 27 / 50 | Improved |
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.93)
- ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification (NeurIPS 2025 · similarity 0.91)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
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.
