Papers2

#Entropy

Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning

Yuanda Xu, Hejian Sang et al.Feb 24arXiv

The paper shows that when training reasoning AIs with reinforcement learning, treating every wrong answer the same makes the AI overconfident in some bad paths and less diverse overall.

#ACE#Reinforcement Learning with Verifiable Rewards#GRPO

Multi-hop Reasoning via Early Knowledge Alignment

Intermediate

Yuxin Wang, Shicheng Fang et al.Dec 23arXiv

This paper adds a tiny but powerful step called Early Knowledge Alignment (EKA) to multi-step retrieval systems so the model takes a quick, smart look at relevant information before it starts planning.

#Retrieval-Augmented Generation#Iterative RAG#Multi-hop Reasoning