Jiajun Wu

Adaptive interleaving of vision and language thoughts in robot planning leads to significant improvements in task execution and reasoning efficiency.

HarmVideoBench reveals that existing benchmarks miss critical layers of harmful video understanding, while a new method boosts model accuracy by over 20%.

State-of-the-art image-to-video models are alarmingly vulnerable to visual prompt attacks, achieving up to 100% success in triggering harmful outputs.

Complex manipulation capabilities can be achieved by dynamically composing simple behaviors, leading to unprecedented precision and adaptability in real-world tasks.

Primitive steerability in VLAs allows for autonomous skill acquisition, enabling robots to learn new tasks without human demonstrations.

A two-loop configuration in LoopCoder-v2 boosts code generation performance by over 50% compared to a non-looped baseline, while more loops actually hinder results.

Naively scaling test-time compute is wasteful; strategically allocating it with DIRECT can enhance embodied agent performance while slashing latency by up to 65%.

VLMs can learn to actively reason and plan in 3D environments by distilling view graphs from self-exploration trajectories, enabling them to surpass even larger models like GPT-4 Pro and Gemini 1.5 Pro on interactive view planning.

Training generative models just got a whole lot easier: GPIC offers 100M permissively licensed, captioned, and safety-filtered images.

MIRA achieves superior mid-training data selection by dynamically constructing source-specific evaluation rubrics, outperforming traditional methods while using half the data.