Qiaozhi He

RRC: Unlocking Generative Reward Models in LLM Reinforcement Learning via Ranking-Based Reward Construction

Chenglong Wang, Ziming Zhu, Yifu Huo, et al.

DaPT: A Dual-Path Framework for Multilingual Multi-hop Question Answering

Yilin Wang, Yuchun Fan, Jiaoyang Li, et al.

This paper addresses the challenge of answering complex multi-step questions across multiple languages using AI systems. The researchers created new multilingual benchmarks for testing these systems and developed DaPT, a framework that processes questions in both the original language and English simultaneously to better retrieve relevant information and generate answers. Their approach significantly outperforms existing systems, achieving 18.3% improvement on the hardest benchmark.