Tencent's WorldCycle: Self-Verifiable RL for Long-Horizon Video World Models
tencent · hf · 2026-08-06
To address compounding errors in long-horizon planning within interactive video world models, Tencent introduced WorldCycle, a self-verifiable reinforcement learning framework.
The key insight is that reversible action cycles enable verification: a sequence combined with its inverse must analytically return to the initial state, providing annotation-free supervision on long-horizon correctness. WorldCycle optimizes two complementary rewards:
- Spatial Closure Reward: Enforces symmetry between mirrored forward and reverse segments.
- Temporal Consistency Reward: Aligns states across repeated cycle executions.
This forces the model to learn actions as consistent state operators rather than memorizing temporal patterns. WorldCycle reduces state returning drift by up to 44% and lifts composite-action accuracy nearly 4x. The team also released CycleBench, a diagnostic benchmark for state-returning abilities.
More from Research
- UC Berkeley Introduces RHI: Optimizing Agent Harnesses to Cut Inference Costs by 60% — ceciletamura · 2026-08-06
- LLMs as Autonomous Cyber Defenders: Multi-Agent Security Research — xuanalogue · 2026-08-06
- Nature Publishes Landmark HCMI: 665 Cancer Organoids from 2,780 Patients Released — anshulkundaje · 2026-08-06
- SKILL-KD: Contrastive Skill Distillation for Weaker LLM Agents — ZhejiangUniversity · 2026-08-06
- Princeton Introduces Skill Entropy to Measure and Boost LLM Cross-Skill Reasoning — princetonu · 2026-08-06
- Tencent Study: VLM Agents Face Severe Safety Risks from Stale Spatial Memory — tencent · 2026-08-06