Post-training quantization for World Action Models

QuantWAMs

Calibrating at the Right Granularity for World Action Models

Aligning every quantization decision with model structure, the joint video–action objective, and reachable closed-loop rollouts.

Jiacheng Zhou,Jinfan Lv,Ruixuan Li,Yan Wang,Longtai Zhang,Wenqiang Zhang,Lizhe Qi

Fudan University · Shanghai Key Lab of Intelligent Information Processing · East China Normal University

Paper arXivCode

WORLD / VIDEO01

PREDICT OBSERVATIONS

JOINT OBJECTIVE

ACTION / CONTROL02

GENERATE ACTIONS

W4A4DOMINANT

29%Peak memory on targeted blocks

1.4–1.6×Block-level speedup

~99.8 %Percentage achieving FP16 precision

2 + 2WAMs / simulation benchmarks

01 / OVERVIEW

Calibration context decides whether quantization transfers.

World Action Models jointly generate future video and actions. Every control cycle denoises iteratively, while early quantization errors alter the states encountered later in the closed loop.

Conventional PTQ methods usually assume open-loop objectives, homogeneous transformers, and static calibration samples. QuantWAMs places each precision decision back into the context it depends on: where evidence is pooled, which state distribution is measured, and which task objective scores it.

context=structure×distribution×objective

The same principle is instantiated with architecture-specific grouping rules for Fast-WAM and LingBot-VA, then evaluated on RoboTwin 2.0, LIBERO, and AgiBot G2 tasks.

02 / METHOD

One principle. Three calibration strategies.

From coordinate-compatible activation evidence to joint video–action gradients and reachable closed-loop states.

QuantWAMs pipeline with outlier calibration, saliency allocation, and real-rollout audit

Shared-basis outlier calibration

Pool channel energy only across coordinate-compatible modules, avoiding false sharing and unstable per-context Top-K masks at small sample counts.

STRUCTURE

Co-training-objective saliency

Allocate weight precision with empirical-Fisher scores from the joint video–action loss, using the more stable layer-level decision granularity.

OBJECTIVE

Fixed-intervention rollout auditing

Replay reachable states from real rollouts to revise denoising-step protection schedules without changing the precision budget.

DISTRIBUTION

INSIDE THE CALIBRATION

See why each granularity matters.

Diagnostics expose the evidence behind channel outliers, layer-level saliency, and trajectory resampling.

Activation outlier evolution before and after QuantWAMs calibration — Pooled-energy Top-K masks remain stable in the low-sample regime.

Layer-level joint video and action saliency map — Joint video–action gradients determine where W8 protection is spent.

Trajectory resampling crossover and recovered energy analysis — V×A IoU matrices. Pairwise top-K channel IoU

03 / RESULTS

Low bit. Closed-loop capability intact.

Success rate, block-level speed, and peak weight-plus-activation memory for the targeted blocks under the W4A4-dominant setting.

Method	Precision	RoboTwin 2.0	LIBERO avg.	Speedup	Memory (GB)
Full Precision	FP16	91.9%	97.6%	1.0×	14.4
GPTQ	W4A16	90.9%	96.0%	1.2–1.3×	5.5
SmoothQuant	W8A8	91.3%	96.4%	1.4–1.5×	7.2
SVDQuant	W4A4	61.1%	73.7%	1.6–1.7×	3.6
SVDQuant*	W4A4	65.9%	75.3%	1.5×	4.2
Atom	W4A4	71.7%	76.2%	1.6×	3.8
Atom*	W4A4	77.2%	82.1%	1.5–1.6×	4.2
QuantWAMs	W4A4	91.7%	97.4%	1.4–1.6×	4.2

* SVDQuant* and Atom* use the same nominal mixed-precision allocation as QuantWAMs. Values are means over three paired protocol seeds; speed and memory are measured on targeted blocks.

RESOURCE PROFILE

About three-tenths of the peak memory.

On Fast-WAM’s targeted video and action blocks, peak weight-plus-activation memory falls from 14.4 GB to 4.2 GB.

FP1614.4 GB

QuantWAMs · W4A44.2 GB

04 / REAL ROBOT

From simulation to AgiBot G2.

QuantWAMs demonstrates deployment feasibility across one-arm pickup and dual-arm coordination tasks, retaining practical manipulation capability at W4A4.

56.7%Average across three tasks

1.4×Target-block speedup

01 · ONE-ARM

Pickup apple

8 / 10

02 · DUAL-ARM

Stack three blocks

5 / 10

03 · DUAL-ARM

Fold a towel

4 / 10

TAKEAWAY

Quantizing a WAM is not only about choosing fewer bits. It is about choosing the right decision granularity with finite calibration data.

STRUCTUREOBJECTIVEDISTRIBUTION

05 / CITATION

Cite QuantWAMs

If this work supports your research, please consider citing it.

@misc{zhou2026quantwamscalibratingrightgranularity,
      title={QuantWAMs: Calibrating at the Right Granularity for World Action Models}, 
      author={Jiacheng Zhou and Jinfan Lv and Ruixuan Li and Longtai Zhang and Yan Wang and Wenqiang Zhang and Lizhe Qi},
      year={2026},
      eprint={2607.28405},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.28405}, 
}