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

Source

• Raw Markdown: paper_intact-2026.md
• PDF: paper_intact-2026.pdf
• Preprint: arXiv 2607.26056v1
• Official project: INTACT project page
• Official code: zju3dv/INTACT-JEPA
• Official checkpoints: INTACT-JEPA/INTACT, pinned paper-e5-goal-v1 revision
• Official author X thread: Junhan Sun / DavidSunok
• Local provenance: papers/intact-2026/x_thread_davidsunok47_2082669880181219830.json, normalized thread Markdown, launch video, and official_artifacts_snapshot.md

Status And Credibility

INTACT is a recent robotics preprint by Junhan Sun, Hao Zhao, and Guofeng Zhang. ArXiv lists version 1 as submitted on 2026-07-28 under cs.RO. No peer-reviewed venue acceptance was verified during ingest, so the results remain author-reported preprint evidence.

The source is credible enough to track as important because it is a direct, current extension of LeWorldModel, exposes the full method and controlled ablations, and has an official project page, MIT-licensed training/evaluation code, audited result records, manifests, and a pinned public checkpoint matrix. The artifact surface is unusually strong for a fresh preprint, but independent reproduction is not yet established: the official release checklist still leaves a clean-clone reference evaluation incomplete, and the frozen paper weights require a compatibility runtime rather than the cleaned current root runtime.

The supplied X post is an official first-author thread. It is useful launch framing, but the pinned paper and artifact snapshots are the source of truth for claims and release status.

Core Claim

A forward action-conditioned world model answers:

Given this state and action, what latent state change follows?

Goal-conditioned deployment asks the inverse query:

Given this state and desired latent change, which action or control-input chunk should realize it?

LeWorldModel and Fast-LeWM answer the second question through CEM search over candidate control-input sequences. INTACT instead trains one conditional action operator on two intent families—an observed physical successor and a future goal—so its conditional mean can act directly. Broad search becomes optional local verification rather than the mandatory control interface.

flowchart LR
  O["observation o_t"] --> E["shared visual encoder E"]
  E --> Z["latent state z_t"]
  N["real successor z_t+1"] --> ML["local intent: z_t+1 - z_t"]
  G["future goal z_g"] --> MG["goal intent: sg(z_g) - z_t"]
  Z --> I["shared INTACT predictor"]
  ML --> I
  MG --> I
  P["previous action context"] --> I
  I --> A["action/control-input chunk distribution"]
  A --> D["Direct mean: 0 candidates"]
  A --> V["optional local CEM verification"]
  Z --> F["forward JEPA predictor + SIGReg"]

Method Contract

INTACT retains LeWM’s encoder, forward latent dynamics, and SIGReg regularization. For a current latent state $z_t$, it constructs two state-conditioned motion-intent coordinates:

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

The realized successor remains attached to the encoder gradient and grounds which physical change followed the demonstrated action. The future goal is stop-gradient because it is a deployment condition, not the next physical observation.

Both calls enter the same predictor with the same four-slot grammar:

$$x_t(m)=\left[z_t,m,z_t\odot m,A(a_{t-1})\right],G_{\eta }(x_t(m))=p_{\eta }(a_t\mid x_t(m)).$$

The action objective is the sum of two proper negative log-likelihoods:

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

The full objective keeps forward latent prediction and SIGReg:

$$\mathcal{L}={\mathcal{L}}_{\mathrm{forward}}+{\lambda }_{\mathrm{SIGReg}}{\mathcal{L}}_{\mathrm{SIGReg}}+{\mathcal{L}}_{I2A}.$$

The local branch alone is inverse dynamics. The goal branch alone is GCSL-like hindsight goal-conditioned imitation. INTACT’s claimed contribution is their end-to-end coupling through one weight-shared action-law operator with asymmetric endpoint gradients—not the invention of either branch objective in isolation.

Conditional action quotient

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

$$y{\sim }_z{y}^{\prime }⟺p_E^{\star }(a\mid z,y)=p_E^{\star }(a\mid z,y^{\prime }).$$

This is an equivalence over the action condition, not pointwise equality of latent endpoints and not a claim that latent dynamics are globally linear. The forward JEPA remains responsible for preserving state needed for future prediction beyond the immediate action.

Evidence And Results

Task-specific one-epoch models

Each task-specific INTACT result averages three independently trained models; each model is evaluated with three 100-episode seeds on the official LeWM protocol.

Task	Direct, 0 candidates	Guarded A, 384 candidates	Published LeWM context
PushT	$85.78\pm 1.54$	$92.22\pm 0.69$	$96.0\pm 4.0$
Cube	$100.00\pm 0.00$	$99.78\pm 0.19$	$74.0\pm 3.0$
Reacher	$97.67\pm 0.00$	$97.44\pm 0.77$	$86.0\pm 5.0$
TwoRoom	$97.89\pm 1.26$	$98.00\pm 1.15$	$87.0\pm 2.5$
Macro	$95.33\pm 0.58$	$96.86\pm 0.38$	85.75

The published LeWM row uses a separate ten-epoch CEM protocol and is landscape context, not a paired significance control. The matched paper result that supports the search claim is narrower: Direct uses zero candidates, while Guarded A centers a $128\times 3$ residual CEM search on the Direct plan and evaluates 384 rather than 9,000 candidate sequences.

Shared four-task encoder

The controlled E5 matrix trains one visual encoder across PushT, Cube, Reacher, and TwoRoom, with task-specific forward and action heads.

Shared-encoder E5 cell	Native inference	Macro SR
LeWM	CEM $300\times 30$	$66.17\pm 2.67$
Goal-intent only	Direct	$80.61$
Goal-displacement INTACT	Direct	$89.39\pm 0.77$
Goal-displacement INTACT, action heads disabled	Pure CEM $300\times 30$	$70.08\pm 1.13$
Goal-displacement INTACT	Guarded A	$90.47\pm 0.84$

The actor-disabled $66.17\to 70.08$ comparison supports a representation-shaping effect beyond merely attaching a policy head. It does not show that the direct head is unnecessary: most of the closed-loop gain appears when the learned intent-to-action interface is used.

Diagnostics

Across 45 eligible E1—E5 checkpoints, predicted—expert action-family kNN overlap correlates with Direct success at $r=0.954$, linear CKA at $r=0.897$, and pointwise action $R^2$ at $r=0.815$. These are mechanism-consistent associations, not independent causal proof; checkpoint family, epoch, objective, and task remain structured covariates. The paper adds a gauge/pair-shuffling intervention, but the general relationship still needs external reproduction and calibrated representation-similarity controls.

Latency

Direct planner-side inference is reported at 2.9—5.5 ms. The approximately $300\times$ claim compares that range with a measured 1.48 s average for actor-initialized CEM $300\times 30$. It is a planner-side comparison, not end-to-end robot/VLA latency, energy, throughput, sensor encoding, environment stepping, or safety-filter latency.

X Thread Claim Audit

Exact author-thread framing	Paper-grounded reading
“LeWM … searches 9,000 sequences to act”	Refers to CEM $300\times 30$ candidate-sequence evaluation in the inherited benchmark protocol.
“Only 1 epoch. 0 search. 95.33% Direct macro SR.”	Correct for the task-specific goal-displacement models. The shared four-task result uses five epochs and reaches 89.39% Direct macro SR.
“2.9—5.5 ms … $\sim 300\times$ lower planning latency”	Correct for the paper’s audited planner-side comparison against 1.48 s actor-initialized CEM; not an end-to-end deployment benchmark.
“Guarded local verification … 23.44× fewer candidates”	Compares 384 candidates with 9,000 and reports 96.86% macro SR. It is local search around the learned Direct plan, not zero search.
“Disable every action head … 66.17% to 70.08%”	Correct for matched shared-encoder E5 actor-disabled pure-CEM evaluation and supports representation shaping.
“Action-family kNN … $r=0.954$; CKA … $r=0.897$”	Correct for the 45-checkpoint pooled diagnostic; correlation is not a quality certificate or external replication.
“Repo (code/models to follow)”	Accurate at thread launch on 2026-07-30 but stale by ingest. Code and a pinned public checkpoint matrix were available by 2026-08-06.

Relationship To Nearby Sources

• LeWorldModel supplies the forward end-to-end JEPA and SIGReg backbone. INTACT adds an amortized deployment-facing action conditional rather than changing only the forward dynamics.
• Fast-LeWM makes CEM candidate evaluation faster through action-prefix-conditioned parallel multi-horizon prediction; INTACT aims to remove broad candidate search from the default path. The two are complementary: Fast-LeWM can accelerate verification, while INTACT can provide the proposal center.
• Sensorimotor World Models uses inverse dynamics as the anti-collapse signal. INTACT retains SIGReg and forward prediction, uses inverse dynamics as one branch, and adds a deployable goal-conditioned branch through the same action operator.
• Inverse Dynamics Regularization remains a necessary boundary. INTACT’s conditional action quotient is identified only on demonstrated support; action aliases, hidden state, delayed effects, policy bias, and action-sufficient partial collapse remain possible.
• Temporal Straightening changes planner-facing latent geometry. INTACT explicitly avoids requiring globally straight latent trajectories and instead amortizes the inverse control query.

Limitations And Gotchas

• The work is an arXiv v1 preprint with no verified peer-reviewed acceptance or independent reproduction.
• Evidence is four simulated visual goal-reaching tasks inherited from LeWM, not real-robot control, stochastic dynamics, numeric multivariate time series, long-horizon open-world tasks, or digital operational systems.
• Three training seeds are a coarse variability estimate.
• The conditional action quotient is learned only on demonstrated state/endpoint support. Unsupported goals or policy-shifted states can make Direct actions unreliable.
• The direct action law is Gaussian. A conditional mean can hide genuinely multimodal control choices, especially around obstacles, homotopy classes, contacts, or irreversible decisions.
• Direct execution amortizes action search but does not make the forward world model unnecessary: fresh-observation replanning and optional verification remain important.
• The action block and short benchmark horizons do not establish long-horizon hierarchical control.
• The $95.33$ result and published LeWM $85.75$ context do not form a paired matched-compute significance comparison.
• The released paper checkpoint matrix has a compatibility boundary: frozen weights use a legacy five-slot Actor grammar, while the current clean runtime uses the paper’s four-slot grammar. Exact reproduction requires the pinned compatibility runtime and adapter.
• The official release checklist has not yet recorded a clean-clone reference evaluation.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Control and counterfactuals	partially closes outside numeric time series	Learns a direct state-and-goal-to-control-input interface jointly with action-conditioned latent dynamics and retains optional model-based verification.	Needs typed interventions, numeric multivariate observations, delayed/failed actions, exogenous variables, constraints, and counterfactual calibration.
Latent-state prediction	adjacent	Keeps a forward JEPA while letting action likelihood shape the shared representation; actor-disabled CEM improves in the matched shared-encoder matrix.	Needs persistent belief state under partial observability, irregular time, long horizons, rare regimes, and dense-state preservation probes.
Dynamic compute and serving efficiency	partially closes outside time series	Replaces mandatory 9,000-candidate CEM with 2.9—5.5 ms Direct inference and exposes 384-candidate local verification as an optional higher-compute path.	Needs end-to-end hardware/energy/throughput accounting, adaptive verifier routing, safety latency, and non-vision deployment evidence.
Anti-collapse and representation quality	adjacent with warning	Couples SIGReg, forward prediction, local inverse likelihood, and goal likelihood; action-family alignment tracks Direct SR better than effective rank in the reported checkpoint set.	Needs calibrated similarity nulls, rare/safety-variable probes, new action repertoires, multimodal inverse heads, and independent reproduction.
Benchmarks and evaluation protocol	warning plus useful controls	Separates Direct, pure CEM, actor-on CEM, guarded verification, task-specific versus shared encoders, action-head-disabled evaluation, seeds, and candidate budgets.	Needs matched-compute external baselines, real systems, OOD factors, stochastic futures, clean-clone reproduction, and end-to-end latency.

For a time-series/world-model transfer, the reusable mechanism is a deployment-supported intent-to-control interface learned jointly with forward latent-state dynamics. A numeric system could map a desired latent-state displacement or target condition to a control-input or intervention chunk, then use the forward model as an optional verifier. The literal visual-control recipe is insufficient when actions are sparse, delayed, failed, constrained, confounded, or effect-equivalent, or when several safe action sequences are valid.

Links Into The Wiki

• INTACT entity
• LeWorldModel
• Fast LeWorldModel
• Sensorimotor World Models
• stable-worldmodel
• Joint Embedding Predictive Architecture
• World Models
• Inverse Dynamics Regularization
• Robotics Time-Series Modeling
• Foundation Time-Series Model Research Agenda
• Contradictions And Open Tensions

Open Questions

• When should Direct execution be trusted, and which uncertainty or OOD score should trigger local verification or broad replanning?
• Can the Gaussian action law be replaced by a calibrated multimodal action-chunk distribution without losing the latency advantage?
• Does the representation-shaping gain survive independent reproduction under matched LeWM data, optimizer, epoch, and solver budgets?
• Can INTACT and Fast-LeWM be combined so a direct proposal supplies the center while action-prefix dynamics verify path constraints in parallel?
• How should the intent coordinate represent variable horizons, delayed effects, irreversible transitions, and partial observability?
• Can the same shared operator map operational target states to typed interventions when actions have status, timing, parameters, failures, and human approval semantics?
• Which probes show that the joint forward/SIGReg/action objective preserves rare safety-critical state outside the demonstrated action law?