About Me

Hi, I am a Postdoctoral Associate at MIT’s Laboratory for Information and Decision Systems (LIDS), mentored by Prof. Cathy Wu. Previously, I was a Research Fellow at Nanyang Technological University (NTU), supervised by Prof. Jie Zhang. I obtained my Ph.D. in Industrial Systems Engineering at National University of Singapore (NUS) in 2024, where I was mentored by Prof. Zhiguang Cao and honored to be advised by Prof. Yeow Meng Chee. I received my B.E. from South China University of Technology (SCUT) in 2019, advised by Prof. Yuejiao Gong.

I am also working closely with NVIDIA, Amazon, Microsoft Research, Symbotic and Grab (SG).

My research is at the intersection of Machine Learning (ML) and Optimization. I have published 30+ papers in top-tier conferences such as ICML, NeurIPS, ICLR, etc, and top-tier journals such as TPAMI, JAIR, TNNLS, TITS, etc. I am an Area Chair (AC) for NeurIPS and actively review for AI conferences and journals, recognized with multiple best reviewer awards.

Welcome to see my publications, fundings, services, experience, and honors & awards. Drop me an email if you’d like to collaborate or discuss with me! You may approach me at:

• Office: 45-611, 51 Vassar St, Cambridge, MA 02139
• E-mail: yiningma [at] mit [dot] edu

💡 Research Interests

My research has focused on the emerging field of “Learning to Optimize (L2Opt)”, where we develop state-of-the-art AI-driven frameworks/approaches for addressing challenging decision-making problems, e.g., combinatorial optimization problems (COPs). My research drives innovative solutions in broad application domains with real-world complexities such as LLM reasoning, LLM for Optimization task formulation, large-scale vehicle routing in logistics, multi-robot coordination in automated warehouse, complex resource scheduling, path planning, etc.

Research Highlights 🔥 - Complete list available here and on Google Scholar

• Neural Combinatorial Optimization (NCO) - Learning end-to-end differentiable solvers, e.g., DACT/NeuOpt (neural k-opt search), N2S/NCS (neural ruin-and-repair search), NeuOpt/PIP (constraint handling), AMDKD (generalization), FER (dynamic embedding), NHDE (multi-objective), MVMoE/SHIELD (multi-task)
• Learning-guided Optimization (LGO) - Accelerating and elevating classical solvers with AI to harness the best of both worlds, e.g., R-DAS (dynamic operator selection), GLEET/ConfigX (dynamic algorithm configuration), L-RHO (long-horizon COP), L2Seg (large-scale VRP), L-RH-PP (automated warehouse robot coordination)
• Knowledge Discovery & Interpretability - Discovering and understanding superhuman decision-making knowledge, e.g., Symbol (rule generation), LLaMoCo (code generation with LLM), Probing NCO (what has NCO learned?)
• Survey, Benchmark & Open-source Library - IET Review (survey of NCO through 2023), MetaBox-v2 (benchmark of learning-guided MetaBBO), RL4CO (library of RL for COP)

Research Keywords

• Machine Learning: Reinforcement Learning, Deep Learning, Large Language Model (LLM), LLM agent, Foundation Model, Multi-Agent Systems, Interpretability
• Optimization: Combinatorial Optimization, Discrete Optimization, Mixed Integer Linear Programming (MILP), GPU-Accelerated Optimization, Black-Box Optimization (BBO)
• Application: Routing, Planning, Scheduling, LLM reasoning, Manufacturing, Logistics

🎉 News

• I'm on the job market! Please contact me if you know of any suitable positions!
• [03/2026] We are hosting the MIT 2026 Summer Research Camp! Welcome to apply!
• [02/2026] Our L2Seg has been accepted as an Oral presentation at ICLR 2026!!
• [01/2026] Three L2Opt papers got accepted by ICLR’26, including 1) L2Seg, accelerating SoTA solvers by 2x to 7x by learning to prun redundant search; 2) CaR, enabling efficient constraint handling via jointly learned feasibility refinement and awareness; and 3) RADAR, promoting asymmetry-aware representations to boost NCO practicality!
• [01/2026] Awarded the MIT Kaufman Teaching Certificate in recognition of dedication to teaching excellence! See the reference letter from the MIT Vice Chancellor!
• [01/2026] One paper on L2Opt got accepted by JAIR, where we introduce L-RH-PP, the first work of NCO for warehouse automation to coordinate multiple robots in Symbotic!
• [01/2026] One paper on L2Opt got accepted by TEVC, where we introduce LLaMoCo, a very first attempt of finetuning LLM for optimization code generation!
• [11/2025] I will organize and serve as a session chair of “AI for Planning and Scheduling” in the International Federation of Operational Research Societies (IFORS), 2026.
• [10/2025] 🎉 Awarded AI Singapore grant “Trustworthy Human-AI Combinatorial Optimization with Language Feedback” (with MIT, NTU, and SMU) see LinkedIn news!
• [09/2025] Three papers on L2Opt got accepted by NeurIPS’25, including 1) the very first paper studying the internal mechanism and intepretability of NCO models (Spotlight); 2) DesignX, a bi-agent learning system tailored for automated algorithm design in black-box optimization; and 3) MetaBox-v2, an up-to-date benchmark platform for MetaBBO!
• [09/2025] We have organized the 1st workshop on Learning-assisted Evolutionary Algorithm Design (LEAD 2025). Watch recordings here; see you all next year!
• [06/2025] Checkout MetaBox-v2🚀! A major update offering broader optimization scenarios (single-objective, multi-objective, multi-modal, multi-task, etc), extensive benchmarks, even faster parallel training/inference, and more support for MetaBBO.
• [05/2025] Our benchmark library got accepted by KDD (Oral) 2025!
• [05/2025] One paper on L2Opt got accepted by ICML’25, where we introduce SHIELD, a novel NCO solver for Multi-Task Multi-Distribution Vehicle Routing Problem (MTMDVRP).
• [04/2025] I will be serving as an Area Chair for NeurIPS 2025!
• [04/2025] Our L-RHO is featured on the Front Page of MIT.edu (see MIT News 🎊)!
• [12/2024] One paper on L2Opt got accepted by ICLR’25, where we introduce L-RHO, leveraging ML to accelerate RHO by up to 54% for Long-Horizon COPs such as FJSP.
• [12/2024] One paper on L2Opt got accepted by AAAI’25 (Oral), where we propose ConfigX, a unified configuration agent that learns to boost diverse evolutionary algorithms.
• [12/2024] We hosted an unofficial L2Opt workshop at NeurIPS 2024 (see photos)!
• [12/2024] ✈️ I will be attending NeurIPS’24 in person. See you in Vancouver, Canada! 🇨🇦
• [11/2024] One paper on L2Opt got accepted by KDD’25, where we propose a neural approach for optimizing diverse, high-quality solutions to Multi-Solution TSP (MSTSP).
• [10/2024] I am selected as NeurIPS’24 Top Reviewer (10%) for both main and DB track!
• [09/2024] One paper on L2Opt got accepted by NeurIPS’24, where we propose Proactive Infeasibility Prevention (PIP) to elevate neural solvers for complex VRP constraints.
• [08/2024] One paper on L2Opt got accepted by TPAMI, where we promoted a neural collaborative framework integrating learning-to-search and learning-to-construct solvers.
• [08/2024] I’m so excited to join Prof. Cathy Wu’s group in MIT!
• [07/2024] I am selected as ICML’24 Best Reviewer (< 3%)!
• [06/2024] Check out - our latest extensive benchmark to unify frameworks and facilitate research in RL-based CO algorithms 🚀.

(Last updated Jan 2026.)