PlayWorld Reveals Quality Gap in High-Scoring World Models
jiqizhixin · x · 2026-08-28
Researchers from HKU, CUHK, Zhejiang University, and Kuaishou Kling introduced PlayWorld, a novel evaluation framework for world models. Instead of fixed action sequences, PlayWorld uses an Agent Player that interacts freely based on scene and long-horizon objectives, assessing geometric consistency, interaction correctness, and physical plausibility. Results show SOTA models like SANA-WM achieve an 80.4% success rate in reaching targets but score only 1.48/5 on quality rubrics, highlighting a significant gap in maintaining spatial structure and object consistency.
More from Models
- GLM 5.3 Flash surges to 4th place in daily OpenRouter usage — Hesamation · 2026-08-28
- Tencent opens Hy4 preview weights: 770B MoE, 49B active, 1M context — Snoo26837 · 2026-08-28
- Similar benchmarks, double the size: Qwen3.8-Flash-Next needs 360GB vs DeepSeek-V4-Flash's 162GB lossless — vini542reddit · 2026-08-28
- Grok 4.6 hits 95% on GPQA Diamond, tied #1 and beating Opus 5 and GPT-5.6 — XFreeze · 2026-08-28
- Hands-on with Tencent Hy4 preview across 8 projects: better frontend taste, stable long-horizon tasks — vista8 · 2026-08-28
- Yoav Goldberg: home robots snapping to "standard" behavior fails everyday users too — yoavgo · 2026-08-28