PhD candidate/ELTE Eötvös Loránd University
Department of Artificial Intelligence/Faculty of Informatics
researching ▍
trajectory separation for scalable credit assignment in cooperative multi-agent reinforcement learning
credit assignment in multi-agent reinforcement learning is especially difficult under shared global rewards. we propose trajectory separation within ppo, a shared encoder processes all observations in a single forward pass while per-entity trajectories isolate individual contributions to the learning signal. in the lux ai environment, our method achieves 14× the resource throughput of global-trajectory variants and improves returns by 240% over ippo and 1,282% over mappo, with only 200k parameters and two hours of training. accepted at iccci 2026.
Nov 2026 → demoBrittle Unlearning
Interactive demo: watch published unlearning methods fail under a fine-tuning attack.
May 2026 → posta scientifically proven way to turn back time
how to organize a lan party at your ai research department, complete with a custom-built tournament manager, 3d-printed trophies, and the inevitable schedule collapse. featuring unreal tournament 2004, counter-strike 1.6, and the worms armageddon we never got to play.
Jan 2026 →