Nojun Kwak

Generative models can transform the landscape of object detection quantization, achieving state-of-the-art performance even in extreme low-bit scenarios.

A novel LLM-based framework reveals how social media timelines can be harnessed for real-time mental health monitoring and analysis.

Naive confidence-based decoding in diffusion LMs can be easily fooled by high-confidence end-of-text tokens and anchor proximity, but a simple suffix-anchored confidence modulation strategy fixes both.

Achieve real-time video depth estimation on edge devices without sacrificing accuracy by ditching full scene reconstruction for targeted geometric refinement.