Jie Wu

SpectraReward reveals that pretrained MLLMs can serve as powerful zero-shot reward models, outperforming traditional methods without the need for fine-tuning or external labels.

SMART achieves a remarkable 0.42 dB improvement in PSNR over state-of-the-art methods, tackling the critical challenge of geometric consistency in light field super-resolution.

Adversaries can exploit overthinking in LVLMs to induce inference delays of up to 6.96x, jeopardizing robotic safety and performance.

Evolving prompts and verifying answers can boost visual reasoning model accuracy by over 19%—a game changer for scaling reliable data in AI.

Federated RAG systems can now be practical: this work achieves a 62x speedup over prior secure methods, while maintaining model utility, by decoupling attention from data localization.

Image editing gets a reasoning upgrade: a chain-of-thought verifier model beats powerful VLMs at judging edits and boosts editing model performance.

LeapAlign unlocks efficient fine-tuning of flow matching models by collapsing long generation trajectories into two-step leaps, enabling direct gradient propagation to early generation steps without prohibitive costs.