Murray Kang

@haoqik322

Ph.D. @UCSanDiego, Research Intern @GoogleDeepMind, B.S. @uwnlp

following
followers

E2S finetuning's claimed gains over two training baselines on new tasks

Its creator says E2S uses GFlowNets to turn expert demonstrations into training data better aligned with a student model.

2 Sources, 14d ago, first seen 14d ago

Uni-LaDiR proposes a shared reasoning space for images, text and 3D point clouds

The introductory post says the framework generates “latent thoughts” with diffusion for VLMs and VLAs.

6 Sources, 13d ago, updated 13d ago

DeliveryGym unveiled for adaptive AI training in simulated Paris

DeliveryGym's creators report a 54.3% improvement in Qwen3-VL-4B's net income during a simulated shift, using earnings as rewards and failures to guide training.

3 Sources, 29d ago, first seen 29d ago

Featured in

E2S finetuning's claimed gains over two training baselines on new tasks

Uni-LaDiR proposes a shared reasoning space for images, text and 3D point clouds

DeliveryGym unveiled for adaptive AI training in simulated Paris

Featured in

E2S finetuning's claimed gains over two training baselines on new tasks

Uni-LaDiR proposes a shared reasoning space for images, text and 3D point clouds

DeliveryGym unveiled for adaptive AI training in simulated Paris