Wenbin Li

Arithmetic errors in LLMs stem from geometric slippages in internal computations, revealing a surprising fragility in their handling of fundamental math.

Task arithmetic works because models internally allocate distinct features to different tasks, and enforcing this specialization via orthogonality regularization unlocks even better editing.

Achieve state-of-the-art performance on ultra-high-resolution remote sensing tasks without the quadratic compute cost, thanks to a query-guided token compression strategy.

MLLMs can aggressively prune visual tokens without sacrificing performance by adapting token reduction strategies to specific classes and prompts.

Forget prompt engineering: this new region proposal network spots objects across diverse datasets without *any* text or image prompts.