Mert Onur Cakiroglu

Achieving a 15% deferral in AI-generated video detection boosts accuracy significantly while slashing compute costs by $7\times$ compared to traditional methods.

Contextual enhancements in time-series forecasting can drastically shift predictive performance, with retrieval methods showing a staggering drop in value across different data structures.

Achieve state-of-the-art image tokenization by learning the quantization geometry itself, outperforming existing methods like FSQ and SimVQ with fewer active codes and lower representation rates.