Heming Zou

TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning

Heming Zou, Qi Wang, Yun Qu, et al.

Enhancing Pretrained Model-based Continual Representation Learning via Guided Random Projection

Ruilin Li, Heming Zou, Xiufeng Yan, et al.

This paper addresses a problem in continual learning (where AI models learn new tasks sequentially without forgetting old ones) by improving how randomly initialized projection layers work with pre-trained models. The authors propose SCL-MGSM, a method that intelligently selects which random features to use rather than using purely random ones, making the system more stable and accurate when learning new tasks with limited examples.