Driving · 2026-11-03

Seeing Through the Illusion: Test-Time Scaling for World Action Models

Cheikh Ahmed, Mahdi Mostajabdaveh, Milad Yazdani, Zirui Zhou

arXiv:2611.03585PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

WISE is a test-time pre-execution selector for frozen world action models. For each decision it samples multiple action–future rollouts, scores each candidate on action validity/temporal regularity, inverse-dynamics consistency between the proposed action and imagined motion, and instruction-conditioned task progress, then executes the highest-scoring action chunk. With four candidates per decision, it improves closed-loop success on RoboLab (Cosmos 3 Edge) and LIBERO (DreamZero).

Why it matters

The authors target a specific failure in world action models: plausible imagined futures may not match the co-generated actions that will be executed. WISE evaluates coupled action–future proposals before acting using only pre-execution signals, aiming to reduce dream-execution divergence by selecting rollouts whose action and imagined motion agree.

Method

  • Sample B paired action–future candidates from a frozen WAM at each step and score each candidate before execution.
  • Compute an action suitability score from interface-based validity checks (including temporal regularity such as velocity/acceleration/jerk when applicable).
  • Compute world-action consistency via embodiment-specific inverse dynamics (IDM) agreement between the WAM’s proposed action and the action recovered from the imagined future; add instruction-conditioned task progress from a frozen task evaluator; sum bounded scores and execute the best candidate. (B=4 in evaluations).

Limitation

WISE depends on auxiliary evaluators: the IDM may misrecover the actions implied by an imagined trajectory, and the progress model may misjudge visually plausible futures.

Abstract (from arXiv)

World action models (WAMs) jointly generate actions and imagined futures, but a plausible prediction need not be realized by its co-generated action. We introduce WISE (\emph{World Action Inference Scaling through Evaluation}), a pre-execution selector that scores each sampled rollout on action validity and temporal regularity, inverse-dynamics consistency between the proposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0\% to 24.0\% on RoboLab with Cosmos~3 Edge and from 75.1\% to 78.3\% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Component ablations and agreement analyses show that the three scores rank candidates differently and that task progress carries the largest weight in the fused decision.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-02 · Indexed by Pub-AI: 2026-11-03 UTC

preprint

Seeing Through the Illusion: Test-Time Scaling for World Action Models

Cheikh Ahmed, Mahdi Mostajabdaveh, Milad Yazdani, Zirui Zhou

Preprint posted: 2026-11-02 · Original record ↗

arxiv
2611.03585
Explicit source versions (1)
  • v1 · 2026-11-02
  1. Preprint available

    Source date: 2026-11-02 · First detected: 2026-11-03 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

World action models (WAMs) jointly generate actions and imagined futures, but a plausible prediction need not be realized by its co-generated action. We introduce WISE (\emph{World Action Inference Scaling through Evaluation}), a pre-execution selector that scores each sampled rollout on action validity and temporal regularity, inverse-dynamics consistency between the proposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0\% to 24.0\% on RoboLab with Cosmos~3 Edge and from 75.1\% to 78.3\% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Component ablations and agreement analyses show that the three scores rank candidates differently and that task progress carries the largest weight in the fused decision.