An Empirical Study of World Model Quantization

TL;DR

The paper investigates the viability of post-training quantization (PTQ) for world models used in visual planning, focusing on long-horizon rollout stability and planning performance. Using DINO-WM as a representative instance, it evaluates multiple PTQ methods under weight-only and weight-activation settings across varying bit-widths, granularities, and horizons up to 50 iterations. The study reveals that quantization effects extend beyond simple accuracy or bit-width trade-offs: group-wise weight quantization can stabilize 4-bit rollouts, activation granularity yields inconsistent benefits, and encoder sensitivity dominates over the predictor, with aggressive low-bit quantization severely misaligning the planning objective with task success. These findings provide practical guidelines for deploying quantized world models under tight computational constraints and highlight distinct, task-dependent failure modes in planning with latent dynamics. The work also underscores the need for planning-aware quantization strategies that account for long-horizon rollout dynamics.

Abstract

World models learn an internal representation of environment dynamics, enabling agents to simulate and reason about future states within a compact latent space for tasks such as planning, prediction, and inference. However, running world models rely on hevay computational cost and memory footprint, making model quantization essential for efficient deployment. To date, the effects of post-training quantization (PTQ) on world models remain largely unexamined. In this work, we present a systematic empirical study of world model quantization using DINO-WM as a representative case, evaluating diverse PTQ methods under both weight-only and joint weight-activation settings. We conduct extensive experiments on different visual planning tasks across a wide range of bit-widths, quantization granularities, and planning horizons up to 50 iterations. Our results show that quantization effects in world models extend beyond standard accuracy and bit-width trade-offs: group-wise weight quantization can stabilize low-bit rollouts, activation quantization granularity yields inconsistent benefits, and quantization sensitivity is highly asymmetric between encoder and predictor modules. Moreover, aggressive low-bit quantization significantly degrades the alignment between the planning objective and task success, leading to failures that cannot be remedied by additional optimization. These findings reveal distinct quantization-induced failure modes in world model-based planning and provide practical guidance for deploying quantized world models under strict computational constraints. The code will be available at https://github.com/huawei-noah/noah-research/tree/master/QuantWM.

Figures (7)

• Figure 1: Activation outliers and scale imbalance in the encoder and predictor of DINO-WM before and after smoothing xiao2023smoothquant, illustrating numerical challenges for low-bit quantization in planning-based world models.
• Figure 2: Open-loop rollouts of the world model under different quantization methods and bit-widths on WALL and Push-T. Given the first frame and a fixed action sequence, the model predicts future observations, which are reconstructed by the decoder. Results are shown for various quantization settings to illustrate their impact on long-horizon prediction quality. The top row shows the full-precision trajectories.
• Figure 3: RTN
• Figure 4: OMSE
• Figure 5: AWQ