PhD student, Mohamed bin Zayed University of Artificial Intelligence
2 papers at NeurIPS 2025
We propose a data-centric residual matching method that significantly improves dataset distillation efficiency and accuracy.
We present Open CaptchaWorld, a benchmark that tests multimodal LLM agents on solving real-world CAPTCHAs via multi-step reasoning and interaction, revealing large gaps between current models and human performance.