INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
Abstract
Forward latent world models predict how actions change a scene, but recover actions for a desired change only through expensive test-time search. We introduce INTACT (INtent-To-ACTion), an end-to-end JEPA that turns action-labeled, reward-free trajectories into a deployable intent-to-action interface. Each transition supplies physical intent z_{t+1}-z_t, while a future goal supplies deployment intent sg(z_g)-z_t. The architecture is isomorphic between the local and goal motion-intent backbone-input graphs through an identical four-slot grammar and shared parameters, and between supported local and goal motion-intent families through action-law semantics induced by the same predictor rather than pointwise latent equality. INTACT also provides intact transfer from RGB evidence to action-effective latent intent coordinates and from intent families to their corresponding action-law families. Asymmetric endpoint gradients ground physical successors and fix future goals as anchors, joining representation learning and control without pointwise latent matching or globally linear dynamics. The resulting coordinates support a robust distributional action law: its conditional mean serves directly as a search-free policy, while sampling remains available for diversity or optional verification. On the four official LeWM tasks, one-epoch, zero-search models reach 85.78\%, 100.00\%, 97.67\%, and 97.89\% success. Optional local CEM centered on the Direct plan reaches 96.86\% macro success using 384 instead of 9,000 candidate sequences, reducing sampling by 23.44times 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, while predicted--expert action-family kNN tracks Direct success at r=0.954. Direct inference takes 2.9--5.5 ms.
Community
[INTACT is an end-to-end unified JEPA that completes the intent-to-action half of LeWM. Instead of using an action-effect world model only for CEM planning, it jointly learns a deployable controller that maps latent motion intent directly to actions. After a single full-data epoch, INTACT achieves 95.33% direct macro SR across four control tasks with zero test-time search and 2.9–5.5 ms inference, reducing planning latency by approximately 300×. Its shared encoder further supports joint four-task training, moving JEPA world models beyond task-specific models toward unified, search-free control.](INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
Forward world models answer “what will happen if I execute this action?”, while control requires the inverse query: “which action realizes this intent?” Existing latent world models typically bridge this gap through expensive test-time action search.
INTACT learns this missing intent-to-action correspondence end to end. A shared conditional operator maps both realized local transitions and deployable goal intents to their induced action distributions. The two intent instances share the same predictor and action supervision, while using deliberately asymmetric gradient paths. This grounds goal-conditioned control in observed physical transitions without forcing pointwise latent alignment or globally linear dynamics.
After only one training epoch, INTACT achieves 95.33% Direct macro success rate across four visual-control tasks, with zero candidate search and 2.9–5.5 ms inference latency. This reduces planning latency by approximately 300× relative to broad CEM control. INTACT also supports joint four-task training with a shared visual encoder and lightweight task-specific heads.
Our analysis further shows that predicted–expert action-family alignment strongly tracks Direct control performance, connecting the proposed conditional action quotient to measurable representation geometry.)
💡 Key ideas
• One operator, two supported intents. The same action-law predictor processes both observed physical displacement and deployable goal displacement.
• End-to-end grounding. Local successor supervision remains attached to the encoder, while the future goal acts as a detached deployment query.
• Conditional action quotient. Intent conditions are related through the expert action laws they induce, rather than through pointwise latent equality.
• Search-free deployment. The conditional mean directly produces an action chunk; small local verification remains optional.
📊 Headline results
• 1 full-data training epoch
• 95.33% Direct macro SR
• 0 test-time candidates
• 2.9–5.5 ms inference
• ~300× lower planning latency
• 96.86% macro SR with only 384 local-verification candidates
• One shared encoder across four visual-control domains
Get this paper in your agent:
hf papers read 2607.26056 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper