Joshua Tian Jin Tee

We show that a single diffusion model can learn from both positive and negative preference pairs through a bounded objective, removing the need for a separate negative model while keeping optimization stable and efficient.

An active learning method for regression that scores candidate samples using the geometry of a dynamic Voronoi tessellation, balancing informativeness, coverage, and representativeness.

This work reframes stepwise preference optimization as an implicit form of gradient-based guidance, offering theoretical insights into its stability and empirical effectiveness.

We distill a multi-step diffusion model into a single-step generator by learning the ODE governing its diffusion trajectory using a PINN-inspired formulation.

We present a novel approach to multimodal dialogue that effectively bridges historical image context to improve the coherence of generated text responses.

Investigates the use of Vector Quantization to defend DRL agents against adversarial attacks, ensuring robust performance in unpredictable environments.

Proposes a calibration method for test-time prompt tuning that uses text feature dispersion to enhance the reliability of vision-language model predictions.

This paper proposes a differentiable calibration metric that eliminates the need for post-hoc tuning, enabling end-to-end optimization of uncertainty estimates.