Source-linked AI summary
World-Coherent Decoding: Self-Verifying Test-Time Planning for World Action Models
Chuhan Zhang, Seiji Ito, Kenta Hoshino, Satoshi Ikehata, Ikuro Sato
TL;DR
World Action Models can generate futures of widely different quality, making candidate selection important for reliable physical execution. WCD ranks multiple future-action candidates with model-internal signals, then uses delayed imagination-reality mismatch to calibrate future selection without updating the backbone; on RoboTwin 2.0, it improves Hard success from 55.80% to 60.90%.
Problem
Stochastic WAM futures can vary from physically consistent and task-relevant to implausible, so the paper studies selection rules for more reliable control.
Method
WCD ranks multiple future-action candidates using flow-based video surprisal and action path effort, then trains a lightweight online predictor from delayed imagination-reality mismatch while keeping the backbone frozen.
Results
WCD improves limited-randomization Hard RoboTwin success from 55.80% to 60.90%, gains 16.43 points on Horizon-3 tasks, and shows qualitative robustness under real Franka visual shifts.
Takeaways & Limitations
The results support selecting reliable futures over merely sampling more candidates as a practical principle for WAM test-time scaling.
Takeaways & Limitations
WCD increases test-time compute through additional parallel candidate generation, although its warmed online predictor adds negligible scoring overhead compared with WAM generation.
Abstract
from arXiv · showhide
World Action Models (WAMs) aim to control robots by stochastically generating visual futures and then decoding actions, but empirical observations indicate that the results can strongly depend on which future is selected. We propose World-Coherent-Decoding (WCD), a self-verifying test-time planning framework that treats WAM rollouts as falsifiable future--action hypotheses. At each decision step, WCD samples multiple candidates from a frozen WAM and ranks them using internal generative signals: flow-based video surprisal for visual plausibility and action path effort for action-generation stability. After execution, the realized observation audits the selected imagination, yielding an imagination--reality mismatch that trains a lightweight online predictor for future candidate selection. Thus, WCD converts delayed self-verification into pre-execution reliability estimation without updating the backbone model. On RoboTwin 2.0, WCD improves Hard success under limited randomized-scene supervision from $55.80\%$ to $60.90\%$, with a $+16.43$ gains on Horizon-3 tasks, and shows qualitative robustness on real Franka visual-shift tests. These results highlight a simple principle: test-time scaling for WAMs depends less on sampling more futures than on selecting reliable ones.
1 Introduction
WAMs can generate materially different future–action hypotheses from the same history, making reliability-aware selection useful before execution. WCD combines internal candidate scoring with delayed self-verification to improve control without updating the frozen backbone.
- 1 Introduction: WAMs jointly model visual futures and action chunks, exposing predicted consequences before physical commitment.The imagined future acts as an implicit goal for decoding the action needed to reach it.
- 1 Introduction: Under identical observation prefixes and instructions, stochastic WAM futures can differ sharply in physical consistency, task relevance, and imagination–reality error.High-error samples may contain nonexistent objects, implausible motion, or gripper-pose drift.
- 1 Introduction: Early imagination–reality mismatch is associated with higher rollout failure rates, making it a behaviorally meaningful reliability signal.The mismatch is revealed only after execution, creating a temporal challenge for pre-execution action selection.
- 1 Introduction: WCD ranks best-of-N future–action candidates with flow-based video surprisal for visual plausibility and action path effort for action-generation stability.The framework is reward-free and verifier-free, using model-internal generative traces.
- 1 Introduction: Delayed mismatch feedback trains a lightweight online predictor that amortizes post-execution self-verification into future pre-execution candidate selection while keeping the WAM frozen.This converts a hindsight audit into a reliability estimate available before later actions are executed.
- 1 Introduction: Under limited-randomization Hard RoboTwin evaluation, WCD improves Hard Avg. by +5.10 points and Horizon-3 success by +16.43 points over the base WAM.The evaluation also includes a real Franka stress test, while ablations attribute gains to structured selection rather than sampling more candidates alone.
2 Related Work
Related work frames test-time computation as a way to improve stochastic generation and robot-policy decoding through candidate selection, refinement, or guidance. WCD addresses the complementary setting of uneven WAM future reliability using internal signals rather than additional evaluators or supervision.
- 2 Related Work: Stochastic-generation systems often require decoding or refinement because sample quality varies substantially under the same condition.Examples include preference or reward models and inference-time refinement for text-to-image generation.
- 2 Related Work: Robot-policy methods guide sampled trajectories with rewards, values, constraints, learned critics, external verification, or action-level filtering.These approaches demonstrate the value of test-time computation but commonly rely on additional supervision or learned evaluators.
- 2 Related Work: WCD targets WAM contexts where future samples are useful but uneven in reliability within the same conditioning history.Its complementary aim is to make test-time imagination more effective under randomized-scene uncertainty.
- 2 Related Work: Figure 2 depicts internal generative signals derived from stochastic conditional-flow sampling, including video surprisal and action path effort.The figure connects candidate generation from a shared cache state with the two signals used for selection.
3 Method
WCD turns a frozen, chunked WAM into a self-verifying best-of-N planner: it ranks imagined future–action candidates before execution, then uses delayed imagination–reality mismatch to calibrate later selection.
- 3.1 Preliminary: At each decision step, the causal WAM conditions on executed interaction history, generates an imagined visual latent future, and decodes its corresponding low-level action chunk.
- 3.3 Delayed Self-Verification: After execution, the realized observation audits the selected imagined future and produces a self-supervised backward error without rewards, success detectors, or external verifiers.
- 3.2 Candidate Selection: WCD samples candidates sharing the same conditioning and ranks them using video surprisal for visual plausibility and action path effort for generation stability.Selection uses only generation-time traces, without rewards, external verifiers, or policy updates.
- 3.2 Candidate Selection: Lower video surprisal favors higher-density imagined futures, while lower action effort favors smoother and more direct action-generation paths.
- 3.4 Online Calibration: WCD trains an online predictor from replayed mismatch labels and generation features, amortizing delayed feedback into a planning-time video score for future candidate ranking.The backbone remains frozen while the predictor and replay buffer persist across episodes of the same task.
- 3.4 Online Calibration: After warm-up, WCD discards surprisal traces and divergence replay, replacing the expensive video score with the lightweight predictor while keeping action effort and the frozen backbone unchanged.
4 Experiments
Experiments evaluate WCD across RoboTwin 2.0 protocols, real-robot visual shifts, and ablations. Results show that delayed self-verification and structured candidate selection improve robustness, especially under limited randomized-scene supervision.
- RoboTwin 2.0 simulation: RoboTwin 2.0 covers 50 bimanual manipulation tasks under standard and limited-randomization training regimes.The limited-randomization protocol reduces randomized-scene supervision while keeping the task set, backbone architecture, and Hard evaluation unchanged.
- Main results: Early imagination–reality mismatch predicts rollout failure: failure rates rise from 4.9% in the lowest-error quartile to 21.2% in the highest.Mismatch is aggregated over the first three executed chunks using task- and depth-normalized error.
- Standard evaluation: 92.28% Hard success versus 91.55% shows WCD improves performance atop a strong frozen WAM, although gains are limited by saturation.Several tasks are already near-perfect under the fully trained standard protocol.
- Limited-randomization and real-robot evaluation: 60.90% versus 55.80% demonstrates WCD’s gain under limited-randomization Hard evaluation, including +16.43 points on Horizon-3 tasks.The backbone remains frozen, and real Franka tests provide qualitative evidence under colored illumination, blur, and cluttered layouts.
- Ablation study: Test-time scaling helps only when additional candidates can be ranked reliably.Random-N fluctuates as candidate count increases, while structured selectors improve more consistently; predictor-only benefits most at N=6–8.
- Ablation study: Ablations show that sampling alone is insufficient: combining video plausibility, action stability, and online prediction reaches the best Hard average.The combined method improves over Random-N by +7.40 and over the strongest single-branch variant by +5.50.
5 Conclusion
WCD treats WAM imagination as falsifiable future–action hypotheses and combines internal candidate scoring with delayed mismatch feedback to select reliable actions without updating the frozen backbone. It improves Hard RoboTwin performance and suggests robustness under visual shifts, while requiring additional test-time computation.
- WCD combines flow-based video surprisal, action path effort, and an online predictor trained from delayed imagination–reality mismatch while keeping the backbone WAM frozen.The framework is reward-free and verifier-free, using delayed self-verification to calibrate candidate reliability.
- Early imagination–reality mismatch is associated with rollout failure, while structured selectors outperform random sampling.These findings support reliability-aware selection rather than relying on stochastic sampling alone.
- Real Franka stress tests suggest that the selection mechanism can recover executable actions under visual shifts.This is presented as a qualitative robustness result.
- WCD improves robustness by spending additional parallel candidate-generation compute, although post-warm online prediction adds negligible scoring overhead compared with WAM generation.The costly flow-surprisal replay is limited to the cold-start stage.