LAQ quantizes robot world action models to 4 bits while protecting predicted futures
Researchers propose latent-aware quantization (LAQ), a 4-bit scheme for world action models that protects the predicted future as well as the action, reporting near-16-bit results on LIBERO-Plus and a real robot arm.
In plain words
World action models are AI systems that predict how a scene will unfold and then choose robot actions based on that prediction. The authors introduce LAQ, a method that shrinks such models to 4-bit numbers so they can run on a robot's onboard computer, while keeping the predicted future accurate because errors there also hurt the actions.
What was announced
The arXiv abstract explains that existing 4-bit quantizers for world action models preserve the generated action but estimate sensitivity one denoising step at a time, leaving the predicted future degraded once the action pathway is protected. LAQ addresses this with a reader metric that measures future error by its effect on the action, and a read count that weights each rounding error by how many denoising steps read it, which the authors say yields a closed-form precision allocation over activations and caches. A gaze-guided refinement step gives attended tokens a second 4-bit pass in the denoising loop and cache, with all weights at 4 bits, no high-precision branch and no per-task recalibration.
On the LIBERO-Plus benchmark, the authors report 76.4% and 83.3% average success on Fast-WAM and Cosmos-Policy, within 1.2 percentage points of 16-bit models and above baselines using 16-bit KV caches. They also report 3.8x model compression, 2.3x lower peak GPU memory for the whole policy, and 1.66x faster running than the 16-bit model. On a real AgileX PiPER arm, LAQ reached 47.0% average success over three tasks, versus 52.0% for the 16-bit model and 30.0% for QuaRot. These figures come from the paper's abstract alone and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract; the full method, experiments and setup details are not covered, and no independent confirmation of the reported results is available.
For security and IT teams
- Scope
- The reported results cover the LIBERO-Plus benchmark and a real AgileX PiPER arm across three tasks, as stated in the abstract.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.