Fugu-MT: arxivの論文翻訳
Daily Report / 2026-06-16

Daily Paper Briefing

2026-06-16 1 Paper

Looped World Models

Authors Hongyuan Adam Lu, Z. L. Victor Wei, Qun Zhang, Jinrui Zeng, Bowen Cao, Lingwei Meng, Mocheng Li, Zezhong Wang, Haonan Yin, Naifu Xue, Minyu Chen, Cenyuan Zhang, Zefan Zhang, Hao Wei, Jiawei Zhou, Haoran Xu, Hao Yang, Ronglai Zuo, Tongda Xu, Yonghao Li, Jian Chen, Hebin Wang, Zeyu Gao, Yang Li, Wei Zhao, Qimin Zhong, Siqi Liu, Yumeng Zhang, Leyan Cui, Zhangyu Wang, Wai Lam
Affiliations FaceMind Research Asia
Categories Method / World Modeling / Iterative latent environment refinement, Method / Transformer Models / Shared parameter transformer blocks, Evaluation / Model Efficiency / Parameter efficiency comparison
License CC BY 4.0

Abstract Overview

This paper introduces Looped World Models (LoopWM), presented as the first application of looped transformer architectures to world modeling. The core idea is to refine latent environment states iteratively using a parameter-shared transformer block, coupled with a spectrally constrained state-retention mechanism intended to keep recurrent updates bounded over long rollouts. The model also supports adaptive early exit, so computation depth can vary with transition difficulty, and a deferred decoding variant that rolls out in latent space and decodes only at the terminal step.

Novelty

The paper’s main novelty is framing world modeling with a looped, recurrent-depth transformer that reuses the same parameters across iterative latent refinement steps, rather than relying on a fixed-depth dynamics model. It also combines this with a spectral stability construction for bounded recurrent dynamics, adaptive early exit at inference, and a deferred decoding scheme for action-conditioned latent rollouts.

Results

On ScienceWorld, the reported overall scores for LoopWM are EM 68.4%, Token F1 85.3%, BLEU-4 80.7%, and Entity 83.9%, compared with 47.2%, 72.8%, 64.4%, and 72.3% for the claude-opus-4-6-max baseline in the paper’s comparison. On AlfWorld, LoopWM reports EM 51.6%, Token F1 80.4%, BLEU-4 71.6%, and Entity 81.1%, giving the best BLEU among the listed models and second-place EM and F1, while entity accuracy appears weaker than some baselines. The deferred decoding analysis further reports gains over external baselines across rollout steps and suggests benefits from accumulating latent computation over longer action sequences.

Key Points

  1. LoopWM uses a parameter-shared looped transformer to iteratively refine latent state transitions, making effective depth less tied to parameter count.
  2. The architecture adds a spectrally constrained recurrent update and adaptive early-exit mechanism to target stable long-horizon rollouts and variable test-time compute.
  3. Empirical results are reported on ScienceWorld and AlfWorld, with especially strong ScienceWorld gains over listed baselines and a deferred-decoding analysis indicating improved multi-step rollout performance.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.