Exact and approximate solution methods for the previously unstudied Precedence-Constrained Quadratic Knapsack Problem (PC-QKP), an NP-hard nonlinear combinatorial optimization problem.

Changkun Guan. "The Precedence-Constrained Quadratic Knapsack Problem." Honors Thesis, Bucknell University, 2024. Advisor: Lucas Waddell.

Crouzeix’s Conjecture, Compressions of Shifts, and Classes of Nilpotent Matrices

Published in Concrete Operators, 2024

Establishes the level set Crouzeix conjecture for several classes of 3×3, 4×4, and 5×5 matrices associated to compressions of the shift, via a geometric analysis of their numerical ranges.

Kelly Bickel, Georgia Corbett, Annie Glenning, Changkun Guan, Martin Vollmayr-Lee. "Crouzeix's conjecture, compressions of shifts, and classes of nilpotent matrices." Concrete Operators, 2024.

Deep Learning-Driven Contextual Stochastic Optimization for Real-Time Order Fulfillment

Published in NeurIPS 2025 Workshop on ML×OR, 2025

A two-stage contextual stochastic optimization framework for real-time e-commerce order fulfillment, with deep neural networks that approximate optimization solutions for deployment at peak scale.

Tinghan Ye, Shuaicheng Tong, Changkun Guan, Beste Basciftci, Pascal Van Hentenryck. "Deep Learning-Driven Contextual Stochastic Optimization for Real-Time Order Fulfillment." NeurIPS 2025 Workshop on ML×OR: Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making, 2025.

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Published in arXiv preprint arXiv:2602.12670, 2026

A community benchmark measuring how structured procedural knowledge packages ("skills") improve LLM agent performance across 87 tasks in 8 domains.

Xiangyi Li, Yimin Liu, Wenbo Chen, et al. (incl. Changkun Guan). "SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks." arXiv preprint arXiv:2602.12670, 2026.

Are Stochastic Multi-objective Bandits Harder than Single-objective Bandits?

Published in arXiv preprint arXiv:2604.07096, 2026

Shows that Pareto regret in stochastic multi-objective bandits scales with the largest objective-wise suboptimality gap, matching single-objective performance — with optimal upper and lower bounds and no dependence on the number of objectives.

Changkun Guan, Mengfan Xu. "Are Stochastic Multi-objective Bandits Harder than Single-objective Bandits?" arXiv preprint arXiv:2604.07096, 2026.

The Proxy Benders Decomposition

Published in arXiv preprint arXiv:2606.07403, 2026

A decomposition framework that replaces Benders subproblem optimization with certified optimization proxies, producing provably valid cuts via a self-supervised predict-project-and-complete mechanism. Up to 161× median speedups on large facility location instances.

Changkun Guan, El Mehdi Er Raqabi, Mathieu Tanneau, Pascal Van Hentenryck. "The Proxy Benders Decomposition." arXiv preprint arXiv:2606.07403, 2026.

Learning Optimization Proxies for Sequential Contextual Stochastic Programs: An Order Fulfillment Application

Published in arXiv preprint arXiv:2606.25362, 2026

Extended version of our NeurIPS 2025 ML×OR workshop paper: neural optimization proxies for real-time order fulfillment that decide ~2,800× faster than solvers while improving fulfillment cost and halving late deliveries.

Tinghan Ye, Shuaicheng Tong, Changkun Guan, Beste Basciftci, Pascal Van Hentenryck. "Learning Optimization Proxies for Sequential Contextual Stochastic Programs: An Order Fulfillment Application." arXiv preprint arXiv:2606.25362, 2026.

End-to-End Supply Chain Planning in the Paper Industry Via Column Generation and Benders Decomposition

Published in arXiv preprint arXiv:2607.16618, 2026

An exact model and hybrid BDCG-DP framework integrating production scheduling, trimming, vehicle loading, and multi-period demand fulfillment in large-scale paper manufacturing — 24.4% cost reduction and 5× faster runtimes on industry instances.

Changkun Guan, Amira Hijazi, Pascal Van Hentenryck. "End-to-End Supply Chain Planning in the Paper Industry Via Column Generation and Benders Decomposition." arXiv preprint arXiv:2607.16618, 2026.

Optimal Combinatorial Testing with Constraints: The Balancing Act

Published in arXiv preprint arXiv:2607.17083, 2026

Revisits pairwise-coverage combinatorial testing on binary parameters, contrasting the unconstrained and NP-hard constrained cases, with the first exact integer-programming algorithm and a fast, often-optimal heuristic.

Thiago Serra, Changkun Guan, Hunter Gehman, Sumit Dhar, Mikey Ferguson, John Hooker, Marcel Schoppers. "Optimal Combinatorial Testing with Constraints: The Balancing Act." arXiv preprint arXiv:2607.17083, 2026.

The Role of Coordinates in Pareto Regret for Adversarial Multi-Objective Bandits

Published in arXiv preprint arXiv:2609.23092, 2026

In adversarial multi-objective bandits, an easier coordinate cannot lower the worst-case order of Pareto regret when the learner does not know which coordinate it is. We prove a minimax lower bound and match it with Poly-INF and an adaptive reward-doubling variant.

Changkun Guan, Mengfan Xu. "The Role of Coordinates in Pareto Regret for Adversarial Multi-Objective Bandits." arXiv preprint arXiv:2609.23092, 2026.

Match the Canvas, Not Just the Answer: Predictability-Aware SFT for Diffusion Language Models

Published in arXiv preprint arXiv:2610.18001, 2026

PACS, a supervised fine-tuning method for diffusion language models that uses the model’s own predictability to better match training to inference, with gains of up to 12.84 percentage points over vanilla SFT on individual reasoning and code benchmarks.

Miao Li, Hanyang Jiang, Hengyu Fu, Baihe Huang, Yating Liu, Yuhang Cai, Changkun Guan, Zhongzhu Chen, Jiecheng Lu, Michael A. Boateng, Pascal Van Hentenryck. "Match the Canvas, Not Just the Answer: Predictability-Aware SFT for Diffusion Language Models." arXiv preprint arXiv:2610.18001, 2026.

OPTAgent: From Standalone Solvers to Deployment-Aware Solver Ecosystems

Published in NeurIPS 2026 Workshop on ML×OR, 2026

An agentic policy-synthesis framework in which an LLM searches compact decision-tree policies that pick optimization methods and configurations, and allocate a shared compute budget across them, for whole collections of problems.

Hanyang Jiang*, Changkun Guan*, Miao Li*, Danlin Li, Sikai Cheng, Pascal Van Hentenryck, Russell Bent. "OPTAgent: From Standalone Solvers to Deployment-Aware Solver Ecosystems." NeurIPS 2026 Second Workshop on ML×OR: Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making, 2026. (*Equal contribution)

talks

Crouzeix’s Conjecture for Nilpotent Matrices

Published: January 05, 2023

Presented our research on Crouzeix’s conjecture for classes of nilpotent matrices at the Joint Mathematics Meetings 2023. The work, joint with Kelly Bickel, Georgia Corbett, Annie Glenning, and Martin Vollmayr-Lee, establishes the level set Crouzeix conjecture for several classes of matrices associated to compressions of the shift and was later published in Concrete Operators.

teaching

Calculus Help Session Leader & Study Group Leader

Peer instruction, Bucknell University, Teaching & Learning Center, 2022–2024

Led weekly calculus help sessions and a dedicated study group for Multivariable Calculus at Bucknell’s Teaching & Learning Center (2022–2024), helping students build problem-solving skills and intuition for multivariable concepts.

Grader, Department of Mathematics & Statistics

Grading, Bucknell University, Department of Mathematics & Statistics, 2022–2024

Graded assignments for Multivariable Calculus, Linear Algebra, and Statistical Inference Theory (2022–2024), providing detailed written feedback on problem sets and proofs.