INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

The Problem: World Models Require Expensive Test-Time Search

Forward latent world models predict how an action changes a scene, but they can only recover the right action for a desired change through expensive search at test time (CEM-style planning over thousands of candidate sequences). Every inference-time query pays the search cost again, and the search is where both latency and failure come from.

Key Idea: The Conditional Mean of an Action Law Is the Policy

INTACT learns an end-to-end JEPA from action-labeled, reward-free trajectories and directly outputs an intent-to-action interface. The architecture is isomorphic between local transitions and future-goal transitions through an identical four-slot grammar with shared parameters, so a single model handles both immediate and goal-directed intents without pointwise latent matching or globally linear dynamics.

The resulting coordinates support a distributional action law whose conditional mean serves directly as a search-free policy, while sampling remains available for diversity or optional verification. Training needs no rewards, only transitions labeled with actions.

Results

• One-epoch, zero-search models reach 85.78%, 100.00%, 97.67%, and 97.89% success on the four official LeWM tasks
• Direct inference takes 2.9-5.5ms
• Optional local CEM centered on the direct plan reaches 96.86% macro success using 384 instead of 9,000 candidate sequences: a 23.44x reduction in sampling while improving pure CEM by 16.00 points
• One shared four-task encoder reaches 89.39% E5 Direct macro and improves every task over jointly trained LeWM
• Predicted vs expert action-family kNN tracks Direct success at r = 0.954, making the learned action law inspectable

What It Means for Developers

This is the purest form of the compute-placement trade: pay at training time so the model itself outputs the action, removing the per-call search step that inference-time scaling methods pay on every query. For teams building plan-and-act agents with fixed latency budgets, the pattern of "train until the conditional mean is the policy, keep sampling only as an optional verification layer" is a reusable design, not just a robotics result - and the 23.44x sampling cut shows where search waste concentrates once a direct plan exists.

Paper: arXiv:2607.26056