← News

Models & toolsSources: arXiv · 20 Nov 2026

LATER world model stores early physical evidence for later action prediction

A new arXiv paper introduces LATER, an action-conditioned world model that keeps fixed memories of early interactions and queries them during planning, reporting gains on a pushing benchmark.

In plain words

Researchers propose LATER, a world model — a system that predicts how a scene evolves under actions — that stores records of early interactions in fixed memory banks and reads from them when predicting the effects of new actions. On a simulated pushing task with distractions, the authors report better success rates and lower prediction errors than a comparable model without this memory design.

What was announced

The paper, posted to arXiv, describes LATER as an action-conditioned world model built on CF-JEPA. It stores completed real transitions in partitioned memories that stay fixed while the model rolls out candidate futures, and a response bridge queries those memories using the current state and a candidate action. Multi-cut auxiliary supervision trains the memory writers, a design the authors link to the distinction between retention and selective readout in working memory research.

On the distracted Two-Stage PushT benchmark, the authors report terminal success rising from 15.5% to 25.0% compared with a reconstructed CF-JEPA. In a separate comparison across six models and two scratch-training seeds, LATER showed the lowest mean physical readout error in both clean and distracted views, cutting along-push mean absolute error under distractors by 20.2% relative to CF-JEPA. Controlled history interventions, where earlier evidence was varied while query observations and actions were held fixed, indicated that later predictions depend on that stored history, with larger effects during pushing than waiting. The figures come from the paper's own experiments and have not been independently verified.

What it does not show

  • The source is the paper's abstract only; full experimental setup, baselines and caveats are not available for verification.
  • Reported results are the authors' own comparisons, including a reconstructed CF-JEPA baseline, and no independent confirmation is provided.

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.