Tag: Reward Design

All the papers with the tag "Reward Design".

• Fight Fire with Fire: Defending Against Malicious RL Fine-Tuning via Reward Neutralization

  grok-3-latest

  Score: 0.66

  Published:2025年5月7日 at  at 17:18

  #LLM, #Reinforcement Learning, #Safety Alignment, #Reward Design, #Fine-Tuning

  本文提出Reward Neutralization框架，通过训练模型生成最小信息拒绝来中和恶意RL微调的奖励信号，显著提升开源模型在攻击下的安全性。

• Reinforced MLLM: A Survey on RL-Based Reasoning in Multimodal Large Language Models

  grok-3-latest

  Score: 0.62

  Published:2025年4月30日 at  at 03:14

  #MLLM, #Reinforcement Learning, #Reasoning, #Cross-Modal Alignment, #Reward Design

  本文系统综述了强化学习（RL）在多模态大语言模型（MLLMs）推理中的应用，分析了算法设计、奖励机制和应用场景，提出了当前局限和未来方向，为多模态推理研究提供了结构化指南。