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

arXiv 2026 37.9 method

TLDR

INTACT learns an end-to-end isomorphic intent-to-action mapping for search-free world models, achieving high success on LeWM tasks.

Reasoning

Strengths include a novel isomorphic architecture that eliminates test-time search and strong empirical results on LeWM benchmarks. Weaknesses are the lack of real-world application and limited discussion of limitations.

Read-first score

Read-first score 37.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=metric,result

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 179.

Keyword Scores

world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Search-free intent-to-action interface: INTACT directly maps intent to action without test-time search, using a distributional action law whose conditional mean serves as a policy.
Isomorphic architecture with a four-slot grammar and shared parameters, making local and goal motion-intent processing structurally identical, and inducing action-law semantics from the same predictor.
Asymmetric endpoint gradients that ground physical successors and anchor future goals, unifying representation learning and control without pointwise latent matching or globally linear dynamics.
Intact transfer from RGB evidence to action-effective latent intent coordinates and from intent families to corresponding action-law families.
One-epoch, zero-search training achieving high success rates on LeWM tasks, with optional local CEM that drastically reduces candidate sampling while improving pure CEM.

Methodology

INTACT is an end-to-end JEPA that learns from action-labeled, reward-free trajectories. It uses a four-slot grammar with shared parameters to process local and goal motion intents isomorphically, and asymmetric endpoint gradients to ground physical successors and fix future goals as anchors. The model produces a distributional action law; the conditional mean is used directly as a search-free policy, with sampling available for diversity or verification.

Key Results

On four official LeWM tasks, one-epoch, zero-search INTACT models achieve 85.78%, 100.00%, 97.67%, and 97.89% success. Optional local CEM centered on the Direct plan achieves 96.86% macro success using only 384 candidate sequences (23.44× reduction vs. 9,000), improving pure CEM by 16.00 points. A shared four-task encoder reaches 89.39% E5 Direct macro, and predicted-expert action-family kNN tracks Direct success at r=0.954. Direct inference takes 2.9–5.5 ms.

Resources

Paper