Sitao Luan

RL Fine-Tuning Heals OOD Forgetting in SFT

Hangzhan Jin

Sicheng Lyu

Mohammad Hamdaqa

The two-stage fine-tuning paradigm of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has empirically shown better reas… (see more)oning performance than one-stage SFT for the post-training of Large Language Models (LLMs). However, the evolution and mechanism behind the synergy of SFT and RL are still under-explored and inconclusive. In our study, we find the well-known claim"SFT memorizes, RL generalizes"is over-simplified, and discover that: (1) OOD performance peaks at the early stage of SFT and then declines (OOD forgetting), the best SFT checkpoint cannot be captured by training/test loss; (2) the subsequent RL stage does not generate fundamentally better OOD capability, instead it plays an \textbf{OOD restoration} role, recovering the lost reasoning ability during SFT; (3) The recovery ability has boundaries, \ie{} \textbf{if SFT trains for too short or too long, RL cannot recover the lost OOD ability;} (4) To uncover the underlying mechanisms behind the forgetting and restoration process, we employ SVD analysis on parameter matrices, manually edit them, and observe their impacts on model performance. Unlike the common belief that the shift of model capacity mainly results from the changes of singular values, we find that they are actually quite stable throughout fine-tuning. Instead, the OOD behavior strongly correlates with the \textbf{rotation of singular vectors}. Our findings re-identify the roles of SFT and RL in the two-stage fine-tuning and discover the rotation of singular vectors as the key mechanism. %reversing the rotations induced by SFT, which shows recovery from forgetting, whereas imposing the SFT parameter directions onto a RL-tuned model results in performance degradation. Code is available at https://github.com/xiaodanguoguo/RL_Heals_SFT

2025-09-08

ArXiv (preprint)

RL Fine-Tuning Heals OOD Forgetting in SFT

Hangzhan Jin

Sicheng Lyu

Mohammad Hamdaqa

The two-stage fine-tuning paradigm of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has empirically shown better reas… (see more)oning performance than one-stage SFT for the post-training of Large Language Models (LLMs). However, the evolution and mechanism behind the synergy of SFT and RL are still under-explored and inconclusive. In our study, we find the well-known claim"SFT memorizes, RL generalizes"is over-simplified, and discover that: (1) OOD performance peaks at the early stage of SFT and then declines (OOD forgetting), the best SFT checkpoint cannot be captured by training/test loss; (2) the subsequent RL stage does not generate fundamentally better OOD capability, instead it plays an \textbf{OOD restoration} role, recovering the lost reasoning ability during SFT; (3) The recovery ability has boundaries, \ie{} \textbf{if SFT trains for too short or too long, RL cannot recover the lost OOD ability;} (4) To uncover the underlying mechanisms behind the forgetting and restoration process, we employ SVD analysis on parameter matrices, manually edit them, and observe their impacts on model performance. Unlike the common belief that the shift of model capacity mainly results from the changes of singular values, we find that they are actually quite stable throughout fine-tuning. Instead, the OOD behavior strongly correlates with the \textbf{rotation of singular vectors}. Our findings re-identify the roles of SFT and RL in the two-stage fine-tuning and discover the rotation of singular vectors as the key mechanism. %reversing the rotations induced by SFT, which shows recovery from forgetting, whereas imposing the SFT parameter directions onto a RL-tuned model results in performance degradation. Code is available at https://github.com/xiaodanguoguo/RL_Heals_SFT

2025-09-08

ArXiv (preprint)

Graph Neural Networks Meet Probabilistic Graphical Models: A Survey

Qian Zhang

2025-04-06

IEEE International Conference on Acoustics, Speech, and Signal Processing (published)

EnzymeFlow: Generating Reaction-specific Enzyme Catalytic Pockets through Flow Matching and Co-Evolutionary Dynamics

Yang Liu

Odin Zhang

Kevin K Yang

Shuangjia Zheng

2024-10-13

NeurIPS.cc/2024/Workshop/AIDrugX (poster)

openreview.net

Reactzyme: A Benchmark for Enzyme-Reaction Prediction

Bozitao Zhong

Liang Hong

Shuangjia Zheng

2024-09-26

NeurIPS.cc/2024/Datasets_and_Benchmarks_Track (poster)

openreview.net

Are Heterophily-Specific GNNs and Homophily Metrics Really Effective? Evaluation Pitfalls and New Benchmarks

Qincheng Lu

Chenqing Hua

Xinyu Wang

Jiaqi Zhu

Xiao-Wen Chang

Guy Wolf

Jian Tang

Over the past decade, Graph Neural Networks (GNNs) have achieved great success on machine learning tasks with relational data. However, rece… (see more)nt studies have found that heterophily can cause significant performance degradation of GNNs, especially on node-level tasks. Numerous heterophilic benchmark datasets have been put forward to validate the efficacy of heterophily-specific GNNs and various homophily metrics have been designed to help people recognize these malignant datasets. Nevertheless, there still exist multiple pitfalls that severely hinder the proper evaluation of new models and metrics. In this paper, we point out three most serious pitfalls: 1) a lack of hyperparameter tuning; 2) insufficient model evaluation on the real challenging heterophilic datasets; 3) missing quantitative evaluation benchmark for homophily metrics on synthetic graphs. To overcome these challenges, we first train and fine-tune baseline models on

2024-09-09

ArXiv (preprint)

The Heterophilic Graph Learning Handbook: Benchmarks, Models, Theoretical Analysis, Applications and Challenges

Chenqing Hua

Qincheng Lu

Liheng Ma

Lirong Wu

Xinyu Wang

Minkai Xu

Xiao-Wen Chang

Doina Precup

Rex Ying

Stan Z. Li

Jian Tang

Guy Wolf

Stefanie Jegelka

Homophily principle, \ie{} nodes with the same labels or similar attributes are more likely to be connected, has been commonly believed to b… (see more)e the main reason for the superiority of Graph Neural Networks (GNNs) over traditional Neural Networks (NNs) on graph-structured data, especially on node-level tasks. However, recent work has identified a non-trivial set of datasets where GNN's performance compared to the NN's is not satisfactory. Heterophily, i.e. low homophily, has been considered the main cause of this empirical observation. People have begun to revisit and re-evaluate most existing graph models, including graph transformer and its variants, in the heterophily scenario across various kinds of graphs, e.g. heterogeneous graphs, temporal graphs and hypergraphs. Moreover, numerous graph-related applications are found to be closely related to the heterophily problem. In the past few years, considerable effort has been devoted to studying and addressing the heterophily issue. In this survey, we provide a comprehensive review of the latest progress on heterophilic graph learning, including an extensive summary of benchmark datasets and evaluation of homophily metrics on synthetic graphs, meticulous classification of the most updated supervised and unsupervised learning methods, thorough digestion of the theoretical analysis on homophily/heterophily, and broad exploration of the heterophily-related applications. Notably, through detailed experiments, we are the first to categorize benchmark heterophilic datasets into three sub-categories: malignant, benign and ambiguous heterophily. Malignant and ambiguous datasets are identified as the real challenging datasets to test the effectiveness of new models on the heterophily challenge. Finally, we propose several challenges and future directions for heterophilic graph representation learning.

2024-07-12

ArXiv (preprint)

The Heterophilic Graph Learning Handbook: Benchmarks, Models, Theoretical Analysis, Applications and Challenges

Chenqing Hua

Qincheng Lu

Liheng Ma

Lirong Wu

Xinyu Wang

Minkai Xu

Xiao-Wen Chang

Doina Precup

Rex Ying

Stan Z. Li

Jian Tang

Guy Wolf

Stefanie Jegelka

Homophily principle, \ie{} nodes with the same labels or similar attributes are more likely to be connected, has been commonly believed to b… (see more)e the main reason for the superiority of Graph Neural Networks (GNNs) over traditional Neural Networks (NNs) on graph-structured data, especially on node-level tasks. However, recent work has identified a non-trivial set of datasets where GNN's performance compared to the NN's is not satisfactory. Heterophily, i.e. low homophily, has been considered the main cause of this empirical observation. People have begun to revisit and re-evaluate most existing graph models, including graph transformer and its variants, in the heterophily scenario across various kinds of graphs, e.g. heterogeneous graphs, temporal graphs and hypergraphs. Moreover, numerous graph-related applications are found to be closely related to the heterophily problem. In the past few years, considerable effort has been devoted to studying and addressing the heterophily issue. In this survey, we provide a comprehensive review of the latest progress on heterophilic graph learning, including an extensive summary of benchmark datasets and evaluation of homophily metrics on synthetic graphs, meticulous classification of the most updated supervised and unsupervised learning methods, thorough digestion of the theoretical analysis on homophily/heterophily, and broad exploration of the heterophily-related applications. Notably, through detailed experiments, we are the first to categorize benchmark heterophilic datasets into three sub-categories: malignant, benign and ambiguous heterophily. Malignant and ambiguous datasets are identified as the real challenging datasets to test the effectiveness of new models on the heterophily challenge. Finally, we propose several challenges and future directions for heterophilic graph representation learning.

2024-07-12

ArXiv (preprint)

Training Matters: Unlocking Potentials of Deeper Graph Convolutional Neural Networks

Mingde Zhao

Xiao-Wen Chang

Doina Precup

2024-02-20

Complex Networks & Their Applications XII (published)