ICML

Junk DNA Hypothesis: Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs "Difficult" Downstream Tasks in LLMs

垃圾DNA假说：修剪小的预训练权重不可逆且单调地损害LLM中的“困难”下游任务

Lu Yin, Ajay Jaiswal, Shiwei Liu, Souvik Kundu, Zhangyang Wang

机构 * University of Surrey ； Eindhoven University of Technology ； University of Texas at Austin ； Intel Labs ； University of Oxford 纠错

AI总结 该研究提出垃圾DNA假说，指出LLM预训练权重中存在关键知识，修剪小权重会单调损害困难下游任务性能，且即使允许持续训练也无法弥补损失。

Comments Published at ICML 2024

R3DM: Enabling Role Discovery and Diversity Through Dynamics Models in Multi-agent Reinforcement Learning

R3DM：通过动态模型在多智能体强化学习中实现角色发现与多样性

Harsh Goel, Mohammad Omama, Behdad Chalaki, Vaishnav Tadiparthi, Ehsan Moradi Pari, Sandeep Chinchali

AI总结 R3DM通过动态模型最大化智能体角色、观察轨迹与预期未来行为间的互信息，提升多智能体协作效率，实验表明其在SMAC和SMACv2环境中显著提升胜率。

Comments 21 pages, To appear in the International Conference of Machine Learning (ICML 2025)