About

Mingyi Hong is an Associate Professor in the Department of Electrical and Computer Engineering at the University of Minnesota Twin Cities. His research focuses on contemporary issues in optimization, information processing, and training, with particular emphasis on foundation models such as language and diffusion models. His work involves addressing challenges in signal processing, wireless communication, and machine learning, contributing to the development of algorithms and methods that enhance the efficiency and effectiveness of these technologies. He holds a Ph.D. from the University of Virginia, obtained in 2011, and a B.Sc. from Zhejiang University in 2005. Mingyi Hong has received numerous honors and awards, including being named an IEEE Fellow in 2025 for his contributions to optimization in signal processing, wireless communication, and machine learning. His scholarly work includes authoring books and publishing extensively in reputable journals, where he has made significant contributions to the fields of optimization algorithms, resource allocation, and signal processing.

Research topics

• Computer Science
• Artificial Intelligence
• Mathematics
• Mathematical optimization
• Algorithm
• Combinatorics
• Distributed computing
• Theoretical computer science
• Applied mathematics
• Pure mathematics
• Geometry
• Mathematical analysis

Selected publications

• HINPool: A Unified Heterogeneous Graph Pooling Framework for Accurate Molecular and Protein Property Prediction

  Proceedings of the AAAI Conference on Artificial Intelligence · 2026-03-14

  articleOpen access1st authorCorresponding

  Graph pooling has gained significant progress in recent years as an effective solution for graph-level property classification tasks. With the emergence of research on Heterogeneous Information Networks (HINs), this paper argues that graph-level datasets for graph classification should be treated as HINs rather than homogeneous graphs to enhance information aggregation. We propose HINPool, a novel and general graph pooling framework for graph-level property classification with HINs. First, we devise a systematic HIN construction procedure from the original data to capture complex interactions. Next, we introduce a type-aware heterogeneous graph pooling method featuring a Type-Aware Selector (TAS) to select essential nodes and a Readout Aggregator (RA) to fuse critical information into a graph-level representation. Finally, a cross-layer fusion function is applied to combine the output embeddings from each graph pooling layer, creating a final graph representation for downstream classification tasks. Our approach achieves near state-of-the-art performance on widely used graph classification benchmark datasets, demonstrating significant improvements in four out of five datasets. This work redefines the strategy for graph-level property classification with HGNNs and heterogeneous graph pooling to model intricate relationships, enhancing performance without requiring extensive domain-specific knowledge.

  PublisherDOI

• MM4Rec: Multi-Source and Multi-Scenario Recommender for Unified User Preference

  Proceedings of the AAAI Conference on Artificial Intelligence · 2026-03-14

  articleOpen access

  As online ecosystems grow increasingly complex, personalized recommendation systems must integrate user preferences across heterogeneous content sources and interaction scenarios. However, conventional methods typically model each source and scenario in isolation, hindering their ability to capture shared and complementary signals across contexts. In this work, we propose MM4Rec, a unified framework for multi-source and multi-scenario recommendation. MM4Rec introduces a Source-Aware Transformer Encoder to jointly model heterogeneous inputs, a Multi-Scenario Behavior Extraction Layer based on a multi-mixture-of-experts architecture to capture scenario-specific dynamics, and a Trend-Aware Learner to enhance temporal representation learning. Extensive experiments on three real-world datasets demonstrate that MM4Rec consistently outperforms strong baselines across standard recommendation metrics. To facilitate future research, we also release two large-scale datasets encompassing diverse sources and scenarios.

  PublisherDOI

• LUME: LLM Unlearning with Multitask Evaluations

  ArXiv.org · 2025-02-20

  preprintOpen access

  Unlearning aims to remove copyrighted, sensitive, or private content from large language models (LLMs) without a full retraining. In this work, we develop a multi-task unlearning benchmark (LUME) which features three tasks: (1) unlearn synthetically generated creative short novels, (2) unlearn synthetic biographies with sensitive information, and (3) unlearn a collection of public biographies. We further release two fine-tuned LLMs of 1B and 7B parameter sizes as the target models. We conduct detailed evaluations of several recently proposed unlearning algorithms and present results on carefully crafted metrics to understand their behavior and limitations.

  PublisherOA PDFDOI

• SemEval-2025 Task 4: Unlearning sensitive content from Large Language Models

  ArXiv.org · 2025-04-02

  preprintOpen access

  We introduce SemEval-2025 Task 4: unlearning sensitive content from Large Language Models (LLMs). The task features 3 subtasks for LLM unlearning spanning different use cases: (1) unlearn long form synthetic creative documents spanning different genres; (2) unlearn short form synthetic biographies containing personally identifiable information (PII), including fake names, phone number, SSN, email and home addresses, and (3) unlearn real documents sampled from the target model's training dataset. We received over 100 submissions from over 30 institutions and we summarize the key techniques and lessons in this paper.

  PublisherOA PDFDOI

• Aligning Large Language Models with Human Feedback: Mathematical Foundations and Algorithm Design

  2025-05-21

  preprintOpen accessSenior author

  This article provides an introduction to the mathematical foundations and algorithmic frameworks used to align Large Language Models (LLMs) with human intentions, preferences, and values. We discuss standard alignment techniques, such as fine-tuning (SFT), reinforcement learning with human feedback (RLHF), and direct preference optimization (DPO). We also explore the theoretical underpinnings of learning from human preferences, drawing connections to inverse reinforcement learning (IRL) and discrete choice models. We present state-of-the-art algorithms in a tutorial style, discuss their advantages and limitations, and offer insights into practical implementation. Our exposition is intended to serve as a comprehensive resource for researchers and practitioners, providing both a foundational understanding of alignment methodologies and a framework for developing more robust and scalable alignment techniques.

  PublisherOA PDFDOI

• MuonBP: Faster Muon via Block-Periodic Orthogonalization

  ArXiv.org · 2025-10-19

  preprintOpen access

  Gradient orthogonalization is a simple strategy that shows great utility in speeding up gradient descent. The Muon optimizer (Jordan, Jin, et al., 2024) combines gradient orthogonalization with first-order momentum and achieves significant improvement in data efficiency over Adam/AdamW (Loshchilov and Hutter, 2019) for language model training. However, when using model parallelism, gradient orthogonalization introduces additional overhead compared to coordinate-wise optimizers (such as AdamW) due to additional gather and scatter operations on gradient matrix shards from different devices. This additional communication can amount to a throughput hit of 5%-10% compared to Adam/AdamW. To remedy this, we propose Muon with Block-Periodic Orthogonalization (MuonBP), which applies orthogonalization independently to matrix shards on each device and periodically performs full orthogonalization to maintain training stability at scale. We show how to adjust the learning rate from the baseline to MuonBP and give convergence guarantees for this algorithm. Crucially, our theory dictates that we use two stepsizes: one for the blockwise orthogonalization steps, and one for the full orthogonalization steps. Our method is simple, requires minimal hyperparameter adjustments, and achieves competitive iteration complexity compared with baseline Muon while providing per-iteration throughput comparable to coordinate-wise methods such as AdamW. When training an 8B model with eight-way tensor parallelism and ZeRO optimizer state sharding, MuonBP achieves 8% throughput increase compared to Muon with no degradation in performance.

  PublisherOA PDFDOI

• Stabilizing Off-Policy Training for Long-Horizon LLM Agent via Turn-Level Importance Sampling and Clipping-Triggered Normalization

  ArXiv.org · 2025-11-25

  preprintOpen accessSenior author

  Reinforcement learning (RL) algorithms such as PPO and GRPO are widely used to train large language models (LLMs) for multi-turn agentic tasks. However, in off-policy training pipelines, these methods often exhibit unstable optimization dynamics and are prone to performance collapse. Through empirical analysis, we identify two fundamental sources of instability in this setting: (1)~a granularity mismatch between token-level policy optimization and turn-structured interactions, and (2) high-variance and unreliable gradient updates induced by off-policy importance sampling and inaccurate advantage estimation. To address these challenges, we propose SORL, \underline{S}tabilizing \underline{O}ff-Policy \underline{R}einforcement \underline{L}earning for Long-Horizon Agent Training. SORL introduces principled mechanisms that align policy optimization with the structure of multi-turn interactions and adaptively suppress unreliable off-policy updates, yielding more conservative and robust learning dynamics. Within this framework, we instantiate two stabilized algorithms: SO-PPO and SO-GRPO. Both algorithms are designed to mitigate gradient variance and prevent optimization collapse without requiring careful early stopping or heuristic tuning. We evaluate SO-PPO and SO-GRPO on a range of multi-turn search benchmarks, including general question answering, multi-hop question answering, and medical multiple-choice QA tasks. Experimental results show that both methods consistently prevent training instabilities and performance collapses observed in standard PPO and GRPO, maintain lower clipping ratios and more stable optimization trajectories, and achieve superior or comparable task performance. These results demonstrate that the proposed algorithm provides a practical, scalable, and general framework for stabilizing reinforcement learning in multi-turn LLM agent training.

  PublisherOA PDFDOI

• A Doubly Stochastically Perturbed Algorithm for Linearly Constrained Bilevel Optimization

  ArXiv.org · 2025-04-06

  preprintOpen accessSenior author

  In this work, we develop analysis and algorithms for a class of (stochastic) bilevel optimization problems whose lower-level (LL) problem is strongly convex and linearly constrained. Most existing approaches for solving such problems rely on unrealistic assumptions or penalty function-based approximate reformulations that are not necessarily equivalent to the original problem. In this work, we develop a stochastic algorithm based on an implicit gradient approach, suitable for data-intensive applications. It is well-known that for the class of problems of interest, the implicit function is nonsmooth. To circumvent this difficulty, we apply a smoothing technique that involves adding small random (linear) perturbations to the LL objective and then taking the expectation of the implicit objective over these perturbations. This approach gives rise to a novel stochastic formulation that ensures the differentiability of the implicit function and leads to the design of a novel and efficient doubly stochastic algorithm. We show that the proposed algorithm converges to an $(ε, \overlineδ)$-Goldstein stationary point of the stochastic objective in $\widetilde{O}(ε^{-4} \overlineδ^{-1})$ iterations. Moreover, under certain additional assumptions, we establish the same convergence guarantee for the algorithm to achieve a $(3ε, \overlineδ + {O}(ε))$-Goldstein stationary point of the original objective. Finally, we perform experiments on adversarial training (AT) tasks to showcase the utility of the proposed algorithm.

  PublisherOA PDFDOI

• RoSTE: An Efficient Quantization-Aware Supervised Fine-Tuning Approach for Large Language Models

  ArXiv.org · 2025-02-13

  preprintOpen accessSenior author

  Supervised fine-tuning is a standard method for adapting pre-trained large language models (LLMs) to downstream tasks. Quantization has been recently studied as a post-training technique for efficient LLM deployment. To obtain quantized fine-tuned LLMs, conventional pipelines would first fine-tune the pre-trained models, followed by post-training quantization. This often yields suboptimal performance as it fails to leverage the synergy between fine-tuning and quantization. To effectively realize low-bit quantization of weights, activations and KV caches in LLMs, we propose an algorithm named Rotated Straight-Through-Estimator (RoSTE), which combines quantization-aware supervised fine-tuning (QA-SFT) with an adaptive rotation strategy that identifies an effective rotation configuration to reduce activation outliers. We provide theoretical insights on RoSTE by analyzing its prediction error when applied to an overparameterized least square quantized training problem. Our findings reveal that the prediction error is directly proportional to the quantization error of the converged weights, which can be effectively managed through an optimized rotation configuration. Experiments on Pythia, Qwen and Llama models of different sizes demonstrate the effectiveness of RoSTE. Compared to existing post-SFT quantization baselines, our method consistently achieves superior performances across various tasks and different LLM architectures. Our code is available at https://github.com/OptimAI-Lab/RoSTE.

  PublisherOA PDFDOI

• Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs

  ArXiv.org · 2025-02-13 · 3 citations

  preprintOpen access

  Large Language Models (LLMs) are increasingly used as chatbots, yet their ability to personalize responses to user preferences remains limited. We introduce PrefEval, a benchmark for evaluating LLMs' ability to infer, memorize and adhere to user preferences in a long-context conversational setting. PrefEval comprises 3,000 manually curated user preference and query pairs spanning 20 topics. PrefEval contains user personalization or preference information in both explicit and implicit forms, and evaluates LLM performance using a generation and a classification task. With PrefEval, we evaluated the aforementioned preference following capabilities of 10 open-source and proprietary LLMs in multi-session conversations with varying context lengths up to 100k tokens. We benchmark with various prompting, iterative feedback, and retrieval-augmented generation methods. Our benchmarking effort reveals that state-of-the-art LLMs face significant challenges in proactively following users' preferences during conversations. In particular, in zero-shot settings, preference following accuracy falls below 10% at merely 10 turns (~3k tokens) across most evaluated models. Even with advanced prompting and retrieval methods, preference following still deteriorates in long-context conversations. Furthermore, we show that fine-tuning on PrefEval significantly improves performance. We believe PrefEval serves as a valuable resource for measuring, understanding, and enhancing LLMs' preference following abilities, paving the way for personalized conversational agents. Our code and dataset are available at https://prefeval.github.io/.

  PublisherOA PDFDOI

Recent grants

• CIF: Small: A Simple and Unifying Optimization Framework for Signal and Information Processing Problems with Min-Max Structures

  NSF · $410k · 2019–2025

• CIF: Small: Collaborative Research: Optimal Provision of Backhaul and Radio Access Networks: A Cross-Network Approach

  NSF · $90k · 2017–2020

• Collaborative Research: MLWiNS: ANN for Interference Limited Wireless Networks

  NSF · $192k · 2020–2025

• Decomposition Framework for Non-convex Nonsmooth Optimization with Applications in Data Analytics

  NSF · $427k · 2017–2023

• CIF: Small: Collaborative Research: Optimal Provision of Backhaul and Radio Access Networks: A Cross-Network Approach

  NSF · $180k · 2015–2018

Frequent coauthors

• Zhi‐Quan Luo

  59 shared

• Meisam Razaviyayn

  43 shared

• Songtao Lu

  42 shared

• Tsung‐Hui Chang

  Shenzhen Research Institute of Big Data

  39 shared

• Xiangyi Chen

  37 shared

• Qingjiang Shi

  32 shared

• Xiao Fu

  Oregon State University

  29 shared

• Alfredo García

  28 shared

Awards & honors

• IEEE Fellow for “contributions to optimization in signal pro…
• IBM Pat Goldberg Memorial Award, honorable mention (2024)
• Pierre-Simon Laplace Early Career Technical Achievement Awar…
• Best Paper Award, IEEE Signal Processing Society (2021, 2022…
• Best Student Paper Award (as advisor), NeurIPS Workshop on S…