Research map / Architectures and efficiency / Quantization

ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models

NeurIPS 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ActQuant, an action-guided mixed-precision post-training weight quantizer for VLA models (HSIC-based per-tensor bit allocation plus Fisher-weighted per-block scale optimization) and OmniModel.cpp, an agentic porting pipeline to a GGML C/C++ runtime.

Problem

Post-training weight quantization of Vision-Language-Action models below 4 bits degrades closed-loop task success sharply, and existing PTQ methods and hardware-unfriendly mixed-precision schemes do not handle this regime well.

Approach

Allocate one bit-width per weight matrix using an HSIC-based action-aware sensitivity score (task relevance minus input redundancy) under a layer-balanced budget, then keep bit-width uniform within each matrix and re-optimize per-block scales using an Action-Mixed Fisher importance weighting.

Key results

• On LIBERO, ActQuant retains 95.0% average success on OpenVLA-OFT at 3.0 bpw and 90.1% at 2.5 bpw (fp16 96.9%), compressing the backbone from 14.3 GB to 2.7 GB.
• On pi0.5, ActQuant achieves 94.8% at 3.0 bpw and 48.0% at 2.0 bpw, where AWQ and GPTQ collapse to 0% at 2.0 bpw.
• On a real UR3 arm, pi0.5 at 3.0 bpw achieves 75.0% average success vs 77.5% fp16 while reducing memory from 6.7 GB to 2.7 GB.

Surprising findings

• QVLA was re-run under a 60-episode calibration budget rather than its original 512 episodes, so its numbers differ from the original paper.
• Baselines AWQ and GPTQ collapse entirely to 0% success at 2.0 bpw on both models.

Limitations

• Weight-only quantization; activations kept at half precision.
• Calibration distribution shift from deployment is not characterized.
• OmniModel.cpp inherits GGML operator coverage; new primitives still require agent reflector iterations.

Experimental setup

Datasets: LIBERO

Benchmarks: LIBERO UR3 real-world tasks

Metrics: success rate bits-per-weight memory (GB) per-token latency (ms)

Baselines: AWQ GPTQ QVLA RTN block quantization

Builds on: QLoRA block quantization GGML / llama.cpp Hilbert-Schmidt Independence Criterion empirical Fisher approximation

Closest papers by meaning

• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.90)
• RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.89)
• Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.89)
• VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.89)