Research
My research lies at the intersection of learning, reasoning, and acting. I develop RL methods and foundation models for complex decision-making, LLM-driven evolutionary systems for automated algorithm discovery, and adaptive agents for multi-agent decision-making and workflow optimization. My long-term goal is to unify RL, language-model reasoning, and agentic optimization into self-improving systems that learn from task feedback and continually refine their own algorithms and workflows.
I am listing some of my research works below.
1. Learning to Optimize
RL4CO: a Unified Reinforcement Learning for Combinatorial Optimization Library
RL4CO is one of the projects I am most proud of and the engineering foundation of much of my learning-based optimization research. Over several years, I gained extensive experience in research infrastructure and baseline development by unifying environments, algorithms, evaluation, and reproducible implementations into a modular ecosystem built on TorchRL, TensorDict, PyTorch Lightning, and Hydra.
RouteFinder: Towards Foundation Models for Vehicle Routing Problems
RouteFinder gave me end-to-end foundation-model experience, from unified task representation and multi-task data generation to distributed pretraining at roughly 200-GPU scale and RL-based post-training. The project also taught me how to balance heterogeneous tasks, evaluate transfer, and adapt a single pretrained model to new routing variants.
USPR: Learning a Unified Solver for Profiled Routing
USPR gave me experience in preference-aligned post-training and unified solver design. We built one model that represents diverse vehicle-client preferences and adapts its decisions through profile-aware embeddings, attention, and score reshaping, aligning a general solver with human-defined priorities without retraining it for every profile distribution.
HADES: Hierarchical Aggregation Deconstruction Search for Vehicle Routing Problems
HADES reflects my experience with hierarchical model design and structure-aware neural search. We modeled within-route, cross-route, and global solution structure through hierarchical representations and tailored positional encodings, improving neural search on large-scale routing problems.
Other works: CAMP, PARCO, RRNCO, etc.
2. Learning to Design Algorithms
ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution
ReEvo is my flagship work on automated algorithm design and a defining step in my research trajectory. As one of the earliest systematic demonstrations of LLM-driven algorithm evolution, it introduced Language Hyper-Heuristics and Reflective Evolution, using performance feedback and LLM reflections to iteratively discover competitive algorithms—and laying the foundation for my work on self-improving agents.
VRPAgent: LLM-Driven Discovery of Heuristic Operators for Vehicle Routing Problems
VRPAgent turns ReEvo's automated-design ideas into an executable optimization system by evolving modular LLM-generated operators inside a correctness-preserving metaheuristic. It gave me end-to-end experience in code generation, validation, genetic search, and rigorous benchmarking; to our knowledge, it was the first LLM-driven approach to outperform strong traditional handcrafted heuristics on established VRP benchmarks.
Other works: PILLM, BuildEvo, TrajEvo, etc.
3. Learning to Improve Agents
JudgeFlow: Agentic Workflow Optimization via Operator Judge
JudgeFlow extends my work from algorithm design to complete LLM-agent workflow optimization. By diagnosing failed executions and assigning operator-level credit, it enables targeted workflow revisions and gave me experience in agent architecture, judge design, failure analysis, and closed-loop optimization.
Other works: EvoNav, etc.
Other: Multi-Agent Robotic RL
I am also working on multi-agent robotic reinforcement learning. The goal is to develop a systematic relational reasoning approach with explicit inference of the underlying dynamically evolving relational structures, and we demonstrate its effectiveness for multi-agent trajectory prediction and social robot navigation.
EvolveHyperGraph: Multi-Agent Trajectory Prediction with Dynamic Relational Reasoning
EvolveHyperGraph represents my early work on multi-agent intelligence. By learning dynamic pairwise and group relations with graphs and hypergraphs, it gave me experience in unsupervised relational representation, trajectory prediction, and interpretable multi-agent modeling.
Multi-Agent Dynamic Relational Reasoning for Social Robot Navigation
This work connects multi-agent reasoning with robotic action by integrating trajectory prediction into a deep-RL social-navigation policy. It gave me end-to-end experience across crowd simulation, multi-agent forecasting, policy learning, safety and social-compliance evaluation, and real-robot validation.
Other: LM Efficiency
I also work on efficient foundation-model systems, spanning dynamic model compression, automated prompting, and hardware-aware optimization. These projects broaden my experience from learning and post-training models to making them practical across inference, deployment, and accelerator design.
POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models
POP reflects my experience in efficient foundation-model inference and structural pruning. We designed a context-aware online pruning method that adapts during autoregressive generation without calibration or retraining, giving me practical experience in latency-aware compression across LLMs, MoEs, and vision-language models.
AoP-SAM: Automation of Prompts for Efficient Segmentation
AoP-SAM reflects my experience in making foundation models practical through lightweight adaptation rather than costly fine-tuning. By learning to generate and refine prompts from frozen SAM embeddings, the project combines prompt automation, test-time adaptive sampling, and efficient vision modeling while preserving zero-shot generalization.
Chiplet Placement and Routing Optimization: A Novel Benchmark and Neural Solver
This work extends my optimization experience from algorithms and models to AI hardware design. We formulated chiplet placement and routing as a hierarchical MDP and developed a neural solver that considers both routing length and electrical performance, giving me experience in hierarchical RL, hardware-aware optimization, and benchmark design.
For more research works, please visit my publications page.
Reinforcement learning is the final step of AI.