How INTACT Eliminates 9,000‑Step Blind Search in World Models by Direct Intent‑to‑Action Mapping

Background

Traditional world‑model control predicts future states but still relies on large‑scale sampling methods such as CEM or MPPI to select the next action. In LeWM benchmarks this means evaluating up to 300×30 = 9,000 candidate sequences per planning step, incurring latency and preventing the model from directly reading action information.

Related Work

LeWM methods train a visual encoder and a forward predictor; the world model acts as a judge for candidate actions generated by CEM/MPPI. Inverse dynamics and goal‑conditioned policies provide either physically grounded actions or direct control, but each suffers from limitations that INTACT aims to unify.

Method

System Overview

INTACT adds a shared conditional action operator (INTACT Predictor) alongside the forward predictor. During training, real local transitions and future goal displacements are fed with identical input structures but different gradient flows, enabling the predictor to learn both local and goal intents. At inference, the goal branch directly generates an action block, while the forward predictor handles future rollout and optional verification.

Two Types of Motion Intent

Local intent derives from adjacent state changes and anchors the model to physically reachable actions. Goal intent comes from the displacement between the current state and the desired future state, available before execution and usable as a direct control condition.

Isomorphic Backbone and Input

Both intents are processed by the same predictor with shared parameters, without forcing their latent representations to be numerically equal. When the intents require similar control behavior, the model maps them to compatible action distributions, while the forward predictor retains information needed for contact, obstacles, and subsequent prediction.

Search‑Free Direct Control with Optional Guarded Verification

INTACT does not eliminate planning entirely; it first produces an intent‑driven plan, executes it, and only resorts to a low‑budget search (e.g., 384 candidates) for verification in complex scenarios.

Loss Design

The training objective combines four components:

World‑model loss: forward predictor predicts latent state after an action; SIGReg maintains a healthy latent distribution.

Local inverse‑dynamics loss: predicts the executed action from adjacent state displacement, embedding physical feasibility.

Goal‑intent action loss: predicts the action from current state to future goal, using stop‑gradient on the goal to avoid treating it as a next physical state.

Joint multi‑task loss: aligns both intent families within a single predictor across four tasks (PushT, Cube, Reacher, TwoRoom).

Results

Single‑Task

Training for one epoch without any search yields an average success rate of 95.33 % across the four LeWM tasks.

Guarded Verification

When local verification is applied, evaluating only 384 candidates raises success to 96.86 %, a 23.44× reduction in samples and reduces planning latency from ~1.48 s to 2.9–5.5 ms.

Multi‑Task

A shared encoder learns all four tasks simultaneously. Goal‑displacement INTACT reaches 89.39 % ± 0.77 % success on the E5 benchmark, 8.78 pp higher than using goal intent alone. Compared with a shared LeWM baseline, improvements are 5.66 pp (PushT), 32.23 pp (Cube), 12.56 pp (Reacher), and 42.44 pp (TwoRoom).

Disabling the INTACT predictor reduces pure CEM success from 66.17 % to 70.08 %, indicating that shared world‑model representations become more plannable, while guarded verification serves as an optional low‑budget validator.

Ablation Studies

Removing the INTACT predictor while keeping encoder, forward predictor, and CEM shows that action supervision improves the shared visual representation, not merely adding a stronger action head. In PushT, pure CEM success rises from 42.2 % to 69.4 % with successive additions of local intent, goal intent, and shared training.

Sharing a single INTACT predictor is crucial: using separate actors for the two intents loses 5.67 pp, and even doubling their capacity cannot match the shared design, confirming that semantic coupling, not parameter count, drives the gains.

Conclusion

INTACT aligns inverse latent representations of intent and action into equivalent classes, enabling a self‑supervised, search‑free control loop while retaining optional verification for difficult cases. The approach reassigns responsibilities: INTACT Direct supplies fast, intent‑driven action plans; the forward predictor maintains future dynamics; search becomes a verification step rather than a direction‑finding process.

Paper: “Isomorphic Intent‑to‑Action Learning for Search‑Free World Models”, arXiv:2607.26056.