Can Test-Time Scaling Improve World Foundation Model?

2025-03-31

AI Summary

Addressing the challenges of prohibitively high pretraining computational costs, limited post-training data, and diminishing returns from model scaling or retraining in World Foundation Models (WFMs), this paper proposes SWIFT—a novel inference-time scaling framework. SWIFT establishes the first empirical validation of an inference-time compute scaling law for WFMs, enabling performance gains without parameter modification or architectural expansion. It introduces a lightweight, scalable inference paradigm integrating rapid tokenization, probability-guided Top-K pruning, and efficient beam search. Evaluated using a custom-built WFM benchmark suite on the COSMOS model, SWIFT demonstrates that optimal inference-time scaling significantly improves prediction accuracy and robustness while maintaining controllable computational overhead. The core contributions are: (1) the first formal confirmation of feasible inference-time scaling for WFMs, and (2) the first efficient, physics-aware inference optimization framework tailored for physical intelligence applications.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Hardware-aware ML
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

World foundation models, which simulate the physical world by predicting future states from current observations and inputs, have become central to many applications in physical intelligence, including autonomous driving and robotics. However, these models require substantial computational resources for pretraining and are further constrained by available data during post-training. As such, scaling computation at test time emerges as both a critical and practical alternative to traditional model enlargement or re-training. In this work, we introduce SWIFT, a test-time scaling framework tailored for WFMs. SWIFT integrates our extensible WFM evaluation toolkit with process-level inference strategies, including fast tokenization, probability-based Top-K pruning, and efficient beam search. Empirical results on the COSMOS model demonstrate that test-time scaling exists even in a compute-optimal way. Our findings reveal that test-time scaling laws hold for WFMs and that SWIFT provides a scalable and effective pathway for improving WFM inference without retraining or increasing model size. The code is available at https://github.com/Mia-Cong/SWIFT.git.

Problem

• Improving world foundation models via test-time scaling
• Reducing computational costs without model retraining
• Enhancing inference efficiency with scalable strategies

Innovation

• Test-time scaling framework SWIFT for WFMs
• Extensible evaluation toolkit with inference strategies
• Compute-optimal scaling without retraining or model enlargement