Shao Tang

Al-phaPO is introduced, a new DAA method that leverages an α -parameter to help change the shape of the reward function beyond the standard log reward, and helps maintain fine-grained control over likelihood displacement and over-optimization.

Quantizing AdamW's second moment in preconditioner space, not state space, slashes validation loss gaps by up to 70% in 4-bit training.

Evaluation-driven development can boost AI agent performance, achieving customer satisfaction levels nearly on par with expert human agents.

Mags-RL lets multimodal LLMs see the forest *and* the trees, using reinforcement learning to guide a super-resolution agent that selectively enhances image regions for improved reasoning without extra annotations.

Training on 500K automatically-curated ophthalmology instructions lets a vision-language model leapfrog general medical models in a specialized domain.

AlphaPO unlocks 7-50% better LLM alignment by showing that reward function *shape* is a surprisingly powerful lever in Direct Alignment Algorithms.