WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models

arXiv 2026 36.8 method

TLDR

WorldCycle uses reversible action cycles to provide self-supervised rewards for RL, reducing drift in long-horizon video world models.

Reasoning

Strengths include a novel self-verification mechanism using action cycles and strong empirical gains (44% drift reduction, 4x accuracy). Weaknesses are the reliance on reversible cycles, which may limit generality, and limited discussion of limitations.

Read-first score

Read-first score 36.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 42%

32.4

Matches configured research keywords against title, abstract, tags, and analysis text. matched=6

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 352.

Keyword Scores

world model

video world model

interactive world model

model-based reinforcement learning world model

generative world model

world simulator

world dynamics prediction

Deep Analysis

Innovations

Reversible action cycles as self-verification: exploiting the principle that a sequence composed with its inverse must analytically return to the initial state, yielding annotation-free supervision on long-horizon correctness.
WorldCycle self-verifiable RL framework that constructs closed action cycles and their repeated executions from ordinary action sequences, and optimizes two complementary rewards: a spatial closure reward enforcing symmetry between mirrored forward and reverse segments, and a temporal consistency reward aligning states across repeated cycle executions.
CycleBench, a diagnostic benchmark for evaluating state-returning ability under complex action structures.

Methodology

WorldCycle constructs closed action cycles (forward and reverse segments) and repeated cycle executions from ordinary action sequences. It then uses reinforcement learning with a spatial closure reward that enforces symmetry between forward and reverse states, and a temporal consistency reward that aligns states across repeated executions, forcing the model to learn actions as consistent state operators and generalizing to out-of-distribution composite cycles.

Key Results

WorldCycle reduces state returning drift by up to 44% and lifts composite-action accuracy nearly 4x over the base model.

Resources

Paper