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

Train a world model to answer the control query it will receive at deployment.

End-to-end JEPA world modeling for goal-conditioned robot control without broad action search.

Junhan Sun1,4  ·  Hao Zhao2,4,†  ·  Guofeng Zhang1,3,†

1State Key Laboratory of CAD&CG, Zhejiang University 2Institute for AI Industry Research (AIR), Tsinghua University 3InSpatio  ·  4RoboParty Lab †Corresponding authors

Why INTACT?  ·  Method Notes  ·  Audited Results  ·  Reproducibility  ·  中文

Project Film

Watch the 61.9-second project film in 1080p  ·  Download the share edition with project QR  ·  Open the interactive project page

A strong representation keeps the information that matters intact. INTACT does exactly that, turning LeWM into a stronger world model.

Why INTACT?

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. The name captures both the structure we impose and
the information we preserve:

• Isomorphic between predictor graphs. Local and goal motion-intent calls
  use the same four-slot input grammar and the same parameters.
• Isomorphic between supported families. Local and goal intent families
  correspond through the action-law semantics induced by that shared
  predictor, not through pointwise latent equality.
• Intact from RGB evidence to latent intent. End-to-end action gradients
  retain action-effective visual information while suppressing nuisance that
  is unrelated to motion intent.
• Intact from intent families to action-law families. The shared predictor
  preserves the supported family correspondence all the way to direct action
  readout.

TL;DR

Forward world models answer "what will happen if I execute this action?"
Yet deployment asks the inverse question: "which action realizes this
intent?" CEM and MPPI answer it by numerically searching over many candidates,
leaving training and inference without a learned semantic correspondence. The
result is often a predictor plus an action searcher, rather than a
self-consistent intent-action model.

INTACT learns that missing correspondence end to end. One conditional
operator maps both observed physical change and deployable goal intent to an
action law. Its conditional mean is a zero-search controller; sampling is
retained only for diversity or optional local verification. No frozen encoder,
extra policy-training stage, or globally linear latent dynamics is required.

1 epoch training  ·  95.33% Direct macro  ·  0 search  ·  2.9-5.5 ms latency

89.39% Shared E5 Direct  ·  96.86% Guarded  ·  23.44x fewer candidates

Core Insight: One Input Grammar, Two Intent Instances

INTACT has one predictor input form. For either intent instance $m_t$, it uses

$$
x_t(m_t)=\big[z_t,m_t,z_t\odot mt,A(a{t-1})\big],
\qquad
G_\eta\left(x_t(mt)\right)
=p\eta(a_t\mid x_t(m_t)).
$$

Only the value and gradient role of $m_t$ change:

$$
mt^{\mathrm{local}}=z{t+1}-z_t,
\qquad
m_t^{\mathrm{goal}}=\mathrm{sg}(z_g)-z_t.
$$

The local instance uses a realized successor to ground which physical
change produced the demonstrated action. The goal instance presents the
same operator with the intent available before acting. Both come from the same
demonstration and are supervised against the same correct $a_t$, but each
supervised conditional remains one triplet $(z_t,m_t,a_t)$ with one endpoint
and one proper NLL:

$$
\mathcal L{\mathrm{I2A}}
=\lambda{\mathrm{local}}[-\log p_\eta(a_t\mid x_t(mt^{\mathrm{local}}))]
+\lambda{\mathrm{goal}}[-\log p_\eta(a_t\mid x_t(m_t^{\mathrm{goal}}))].
$$

There is no direct loss between the two endpoints or their latent
displacements. Instead, the shared likelihood creates a conditional action
quotient: at a fixed state, intents are equivalent when they induce the same
expert action law. Under a task-appropriate tolerance, nearby predictions
$\hat a_t^{(1)}$, $\hat a_t^{(2)}$, and the demonstrated $a_t$ can therefore
belong to the same action-equivalence neighborhood. This distributional view
makes direct control less sensitive to small prediction errors and helps limit
closed-loop drift, while the forward JEPA keeps the richer world information
needed for prediction.

Method

The physical successor remains attached to ground reachability; the future
goal is a stop-gradient deployment anchor. INTACT aligns the two supported
condition families through the actions they induce, without matching their
endpoints, imposing globally linear latent dynamics, freezing the encoder, or
adding a phase-2 controller. The goal likelihood alone is a goal-conditioned
imitation objective; full INTACT is its shared, end-to-end coupling with the
attached physical likelihood and forward JEPA.

INTACT retains the forward JEPA and adds one action-law predictor with a matched
input grammar for physical and deployable intents. The key construction is not
two unrelated auxiliary losses: both calls share parameters and a proper action
likelihood, while their upstream gradient routes remain deliberately asymmetric.

The four-domain model shares one visual encoder and keeps lightweight,
task-specific forward/action heads:

See Method Notes for the statistical construction, gradient
contract, and the distinction from inverse dynamics, goal-conditioned behavior
cloning, and post-hoc sampling priors.

Results

Task-Specific, One Epoch

The matched task-specific setting trains three models per task. Each checkpoint
is evaluated with three seeds and 100 episodes per seed on the official LeWM
protocol.

One epoch of goal-displacement INTACT reaches 95.33 +/- 0.58% Direct macro
SR with no candidate search. The same checkpoints reach 96.86 +/- 0.38%
with 384-sequence local verification. Published LeWM numbers use its separate
10-epoch CEM protocol and serve only as landscape context, not paired
significance controls. Exact task values and the full matched inference matrix
remain available in Audited Results.

One Shared Encoder

At epoch 5, Goal-displacement INTACT reaches 89.39 +/- 0.77% Direct macro SR
with one encoder shared across all four visual domains. The matched shared LeWM
baseline reaches 66.17 +/- 2.67% with CEM 300x30. With all INTACT action heads
disabled, pure-CEM macro still rises to 70.08 +/- 1.13%, separating
representation shaping from direct action readout.

Theory Meets Measurement

At fixed state $z$, INTACT treats two endpoint conditions as equivalent when
they induce the same expert action law:

$$
y\sim_z y' \iff p_E^\star(a\mid z,y)=p_E^\star(a\mid z,y').
$$

This conditional action quotient predicts that control should track the
relation between predicted and expert action-law families, not task clustering
or latent rank alone. Across 45 eligible shared-encoder checkpoints,
predicted-expert kNN overlap correlates with Direct SR at r = 0.954 and
linear CKA at r = 0.897; pointwise action $R^2$ is weaker at r = 0.815.

Paper, Website, and Code

This repository is currently a public research preview. The method record,
audited result tables, attribution, and reproducibility contract are organized
here first. Training/evaluation source, anonymous submission artifacts, and
checkpoints will be released in stages after their licensing, provenance, and
double-blind review checks are complete.

The detailed release gates are tracked in Release Plan.

Citation

An arXiv preprint is available at arXiv:2607.26056.
Please prefer the machine-readable CITATION.cff record.

@misc{sun2026intact,
  title         = {INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models},
  author        = {Sun, Junhan and Zhao, Hao and Zhang, Guofeng},
  year          = {2026},
  eprint        = {2607.26056},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.26056}
}

Acknowledgements

INTACT builds on the LeWM and stable-worldmodel research ecosystem. Evaluation
is reported on the official LeWM protocol and, where explicitly marked, the
separate CLEAR-LeWM evaluator.
See NOTICE.md for provenance and licensing boundaries.

Questions: [email protected]