IEEE

Reward Hacking in Reinforcement Learning and RLHF: A Multidisciplinary Examination of Vulnerabilities, Mitigation Strategies, and Alignment Challenges

Abstract: Reinforcement Learning (RL) agents optimize policies based on provided rewards, yet may exploit unintended loopholes in the reward design, a phenomenon known as reward hacking. With the rise ...
IEEE

Learning from Failures: Translation of Natural Language Requirements into Linear Temporal Logic with Large Language Models

Abstract: Formalization of intended requirements is indispensable when using formal methods in software development. However, translating Natural Language (NL) requirements into formal specifications, ...