TRAD: Enhancing LLM Agents with Step-Wise Thought Retrieval and Aligned Decision
In SIGIR 2024
This paper proposes TRAD, a novel LLM agent achieving step-wise relevant demonstration selection via Thought Retrieval and Aligned Decision.
Publications
Is Risk-Sensitive Reinforcement Learning Properly Resolved?
In arXiv 2023
This paper provides an in-depth analysis on the biased optimization issue of existing risk-sensitive reinforcement learning (RSRL) methods, and proposes Trajectory Q-Learning (TQL), a novel RSRL framework that is proven to learn the optimal policy w.r.t. various risk measures.
Learning Enhanced Representations for Tabular Data via Neighborhood Propagation
In NeurIPS 2022
This paper proposes PET, a novel architecture that Propagates and Enhances the Tabular data representations based on the hypergraph for target label prediction.