Alex Open Research Wiki

INTACT

3 min read

INTACT

Summary

INTACT (INtent-To-ACTion) is an end-to-end JEPA world model that learns a direct intent-to-action interface alongside forward latent dynamics. It maps either an observed physical latent change or a desired goal latent change to an action/control-input chunk through one shared conditional operator.

current latent state + desired latent change + previous-action context
  -> shared conditional action operator
  -> action-chunk distribution
  -> Direct mean with zero candidates
  -> optional local model-based verification

The method is introduced in INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models.

Interface

INTACT uses two intent instances:

Both enter the same four-slot predictor grammar:

The local call is an inverse-dynamics likelihood with an attached physical successor. The goal call is a GCSL-like hindsight imitation likelihood with a stop-gradient deployment goal. The full method couples both calls to the same predictor while retaining the LeWorldModel forward predictor and SIGReg objective.

Evidence Boundary

The strongest reported task-specific result is % macro success after one full-data epoch with zero candidate search on the four official LeWM tasks. Planner-side Direct latency is 2.9—5.5 ms. A 384-candidate local verifier reaches % macro success. A separate shared-encoder E5 matrix reaches % Direct macro success and shows an actor-disabled pure-CEM gain from % for matched LeWM to %.

These are author-reported simulated visual-control results from an arXiv v1 preprint. They do not establish real-robot transfer, stochastic or multimodal control, long-horizon planning, numeric multivariate time-series control, or independent reproduction. The approximately latency claim is planner-side, not end-to-end deployment latency.

Official Artifacts

The public release includes MIT-licensed training/evaluation code, audited results, checkpoint manifests, and an immutable checkpoint revision. The frozen paper checkpoints require the release’s pinned compatibility runtime because their Actor grammar differs from the cleaned current runtime. The official release checklist had not yet recorded a clean-clone reference evaluation at ingest time.

Role In The Wiki

INTACT is the amortized intent-to-control branch of the LeWorldModel line:

  • LeWorldModel supplies stable end-to-end JEPA latent dynamics.
  • Fast-LeWM accelerates candidate-sequence evaluation but retains CEM as the controller.
  • INTACT learns a Direct action proposal so broad search is no longer mandatory; the forward model remains useful for optional verification and recurrent replanning.
  • Sensorimotor World Models uses inverse dynamics as the anti-collapse objective; INTACT combines a local inverse branch with a deployable goal branch while retaining SIGReg and forward prediction.

Relation To Foundation TSFM Agenda

Use the source-level slot verdicts in INTACT. The transferable object is a deployment-supported map from maintained latent state plus desired state change to an action/control-input or intervention chunk, with forward latent dynamics retained as a verifier. A time-series version must model partial observability, irregular time, delayed and failed actions, typed constraints, exogenous variables, multimodal control choices, and rare safety-critical state.

Graph View

Backlinks