EWMBench

Headlines · 3

1. 4d
   Qiyuan Robotics releases PrimeSim robot world model

   Qiyuan Robotics’ Embodied Intelligence Research Center released PrimeSim, an interactive world-model simulator for robotic manipulation, according to 36Kr and Machine Heart. Built from a pretrained video-generation model, PrimeSim reportedly supports cross-embodiment few-shot transfer, continuous rollouts exceeding two minutes, and reinforcement learning; its weights and inference and training code are planned for open-source release. The research team reports a 6.09 composite score on EWMBench and first place in three of WorldSimProbe’s five dimensions, while 36Kr says reinforcement learning raised policy success from 61.7% to 73.2%.

   机器之心 · 🔥 0
2. Oct 16
   UBTech says Thinker-WM tops three embodied world-model benchmarks

   UBTech says its in-house unified multimodal world model Thinker-WM ranked first on three embodied world-model benchmarks: WorldOlympiad (overall score 0.740), EWMBench (4.66 total) and WorldArena 1.0 Track-1, claiming it beat models including Qwen-RobotWorld and Sora 2. The model is built around a Unified Dynamics Transformer that encodes RGB video, depth and robot actions into one shared representation space, letting a single set of weights switch between conditioning and generation modes for six capabilities such as joint video-action generation, world prediction and policy reasoning. UBTech says training data comprised roughly 20,000 hours of first-person human video, 10,000 hours of real robot data and 1,000 hours of simulation, and that the model has been validated on its own robot platforms; all figures come from UBTech-supplied material republished with authorization by QbitAI.

   量子位（原生 RSS） · 🔥 0
3. Oct 15
   EVEWorld cuts world-model object-count anomalies by 85.7% on DreamGenBench

   Researchers at China Merchants Group's Advanced Technology Research Institute and Nanyang Technological University propose EVEWorld, a physical-evolution supervision framework that uses instance-guided recovery (IGR) and temporal instance alignment (TIA) to keep manipulated objects consistent within and across generated video frames. The authors report that on DreamGenBench, EVEWorld cut the model-laziness rate (MLR) from 11.11% to 1.59%, an 85.7% relative reduction versus plain SFT, while overall Gemini-IF rose 13.6% relative; on WorldArena 2.0 Track 1, a FlowWAM-based SupervisionWM scored 70.17. These are author-reported results without independent replication, and on the FlowWAM backbone TIA alone reached a 22.54% MLR, better than the full combination's 35.21%.

   36氪 人工智能 · 🔥 0

Experience and discussion from the community

Share my EWMBench experienceAsk about EWMBench