Runyu Lu Photo

Hello! I’m Runyu Lu (陆润宇), a Ph.D. Candidate specializing in Game Theory and Reinforcement Learning at the University of Chinese Academy of Sciences (UCAS), supervised by Prof. Dongbin Zhao.

My research focuses on equilibrium learning, aiming to develop model-based, offline, or online Multi-Agent Reinforcement Learning (MARL) algorithms with last-iterate convergence guarantees in dynamic games or multi-agent systems. I also proposed the Equilibrium Policy Generalization paradigm for real-time robust strategies under time-varying game dynamics.

Prior to my Ph.D., I earned my B.Eng. in Computer Science and Technology from the University of Chinese Academy of Sciences.

📰 News

  • [Apr 2026] I am attending the 14h International Conference on Learning Representations in Rio de Janeiro, Brazil. My poster will be presented on Thu, Apr 23, 2026 at Pavilion 4 P4-#4705 (3:15 PM – 5:45 PM). Come to chat with me if you will be there!

  • [Jan 2026] Our paper “R2PS: Worst-Case Robust Real-Time Pursuit Strategies under Partial Observability” was accepted to ICLR 2026!

  • [Nov 2025] I attended the 7th International Conference in Distributed Artificial Intelligence (DAI 2025) at King’s College London, UK.
    • Organized the workshop “LLMs in Games: Reasoning, Strategy, and Distributed Intelligence” (Nov 21)
    • Gave a talk on “Divergence-Regularized Reinforcement Learning for Equilibrium Policies in Adversarial Games” for the workshop.
    • Made an oral presentation for our accepted paper, “ARAC: Adaptive Regularized Multi-Agent Soft Actor-Critic in Graph-Structured Adversarial Games” (Nov 24).
  • [Sep 2025] Our paper “Equilibrium Policy Generalization: A Reinforcement Learning Framework for Cross-Graph Zero-Shot Generalization in Pursuit-Evasion Games” was accepted to NeurIPS 2025!

📚 Selected Publications

Conference Papers

  • [NeurIPS 2025] Equilibrium Policy Generalization: A Reinforcement Learning Framework for Cross-Graph Zero-Shot Generalization in Pursuit-Evasion Games.

    Summary: Proposes the Equilibrium Policy Generalization (EPG) framework, aiming to find robust policy representation across changing game dynamics. Under pursuit-evasion games (PEGs), EPG enables cross-graph zero-shot generalization of policies by training against the opponents generated from a near-optimal equilibrium oracle. Establishing the first two generalizability properties in this domain. Designing a grouping mechanism and a sequence model for scalability.

  • [ICML 2025] Constrained Exploitability Descent: An Offline Reinforcement Learning Method for Finding Mixed-Strategy Nash Equilibrium.

    Summary: Introduces Constrained Exploitability Descent (CED), a model-free offline RL algorithm for adversarial Markov games (MGs). Theoretically proves that CED converges to a stationary point where the min-player policy exhibits the mixed-strategy Nash equilibrium property. CED significantly outperforms existing pessimism-based and self-play methods in offline equilibrium finding.

  • [ICLR 2025] Divergence-Regularized Discounted Aggregation: Equilibrium Finding in Multiplayer Partially Observable Stochastic Games.

    Summary: Presents Divergence-Regularized Discounted Aggregation (DRDA), a multi-round learning system for Multiplayer Partially Observable Stochastic Games (POSGs). Extends discounted FTRL theory to POSGs, proving last-iterate convergence to Nash distribution in a single round, and convergence to an exact Nash Equilibrium (NE) over multiple rounds.

Journal Paper

  • [IEEE TNNLS 2024] Last-Iterate Convergence to Approximate Nash Equilibria in Multiplayer Imperfect Information Games.

    Summary: Constructs the continuous-time dynamic Imperfect-information Exponential-decay Score-based Learning (IESL) for multiplayer imperfect information games (IIGs). Proves last-iterate convergence to approximate Nash equilibria, which is highly beneficial for deep RL implementations. Empirically verified to outperform CFR and RD variants in multiplayer poker.

Working Paper

  • [ICLR 2026] R2PS: Worst-Case Robust Real-Time Pursuit Strategies under Partial Observability.

    Summary: Proposes Worst-Case Robust Real-Time Pursuit Strategies (R2PS) for graph-based PEGs under partial observability and asynchronous moves. Integrates a belief preservation mechanism with the EPG framework to learn robust, real-time GNN policies against asynchronous-move evasion strategies, achieving superior zero-shot generalization.