[1] Eigenspectrum Analysis of Neural Networks without Aspect Ratio Bias

Yuanzhe Hu, Kinshuk Goel, Vlad Killiakov, Yaoqing Yang

ICML 2025

Short Summary: FARMS, a method for normalizing weight matrices by subsampling with a fixed aspect ratio based on M-P Law, improves the accuracy of eigenspectrum analysis and layer-wise hyperparameter assignment across various domains, including image classification, scientific machine learning, and large language model pruning.

Code | Paper and Review | ICML Website |

[2] Model Balancing Helps Low-data Training and Fine-tuning

{Zihang Liu#, Yuanzhe Hu#}, Tianyu Pang, Yefan Zhou, Pu Ren, Yaoqing Yang

EMNLP 2024 , Oral (168/6105=2.75%), Meta Review OA=5.0

Short Summary: This paper introduce a modified layer-wise learning rate scheduler, improves low-data training and fine-tuning performance in both NLP and SciML by balancing training quality across model layers.

Code | Paper | Video | Review |

[3] Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions

{Yuanzhe Hu#, Yu Wang#}, Julian McAuley

Under Review / ICML 2025 LCFM Workshop

Short Summary: MemoryAgentBench is a new benchmark designed to evaluate four core competencies of memory agents in Large Language Models, highlighting the need for improved memory mechanisms.

HF Dataset | Code | Paper |

[4] M+: Extending MemoryLLM with Scalable Long-Term Memory

Yu Wang, Dmitry Krotov, Yuanzhe Hu, Yifan Gao, Wangchunshu Zhou, Julian McAuley, Dan Gutfreund, Rogerio Feris, Zexue He

ICML 2025

Short Summary: M+ enhances long-term information retention in large language models by integrating a retriever-based long-term memory mechanism, outperforming MemoryLLM and other baselines in knowledge retention tasks.

机器之心 | HF Model | Code | Paper | Review |