DynThinker
A framework that unifies static 3D and dynamic 4D spatial reasoning. Through two components, Decoupled Exponential Moving Average (D-EMA) and Spatial-Dynamic Fusion (SDF), it mitigates spatial-dynamic interference and improves the spatial intelligence of vision-language models.
Timeline (last 90 days)
DynThinker is a unified framework aimed at jointly strengthening static and dynamic spatial reasoning, whose core consists of two complementary components, D-EMA and SDF.
DynThinker was evaluated on ten metrics covering nine static 3D and dynamic 4D benchmarks, achieving the highest average score among all compared models.
DynThinker reaches 56.0 on ReVSI and 67.2 on DSR-Bench, and achieves 51.2 and 40.1 respectively under the per-sample and grouped evaluation protocols of DSI-Bench.
DynThinker demonstrates that static 3D and dynamic 4D capabilities can simultaneously reach leading levels in the same model, rather than necessarily trading one for the other.
DynThinker's current performance is still established on its chosen benchmarks, and its generalization ability in more open and more complex real-world scenarios remains to be further verified.
All Facts (5)
DynThinker is a unified framework aimed at jointly strengthening static and dynamic spatial reasoning, whose core consists of two complementary components, D-EMA and SDF.
50%UnverifiedDynThinker was evaluated on ten metrics covering nine static 3D and dynamic 4D benchmarks, achieving the highest average score among all compared models.
50%UnverifiedDynThinker reaches 56.0 on ReVSI and 67.2 on DSR-Bench, and achieves 51.2 and 40.1 respectively under the per-sample and grouped evaluation protocols of DSI-Bench.
50%UnverifiedDynThinker demonstrates that static 3D and dynamic 4D capabilities can simultaneously reach leading levels in the same model, rather than necessarily trading one for the other.
50%UnverifiedDynThinker's current performance is still established on its chosen benchmarks, and its generalization ability in more open and more complex real-world scenarios remains to be further verified.
50%