Hanmin Li (李 瀚民)

I am a Ph.D. candidate in Computer Science at KAUST working on optimization, distributed training, scalable learning systems, efficient LLM training, agentic AI, LLM post-training and model compression.

June 2026 – Present

My work focuses on long-horizon LLM agents, Mixture-of-Experts (MoE) model pruning, activation steering, and post-training.

Long-Horizon Agent Platform

  • Architected an end-to-end platform for DeepWork long-horizon tool-use agents on the GDPval benchmark.
  • Served Qwen3.6-35B-A3B on nodes with eight NVIDIA A100 GPUs each and integrated a customized agent harness, trajectory capture, artifact validation, and LLM-as-a-judge evaluation.

MoE Observability and Pruning

  • Built exact token-level MoE observability across complete agent trajectories and converted routing statistics into deployable pruned checkpoints.
  • Reduced routed-expert banks from 256 to 160 experts per layer, eliminating approximately 12.1 billion parameters and reducing BF16 model-weight memory by 34%, while preserving shared experts, top-k execution, and benchmark performance.

Expert Specialization

  • Identified and quantified task-dependent expert specialization across industry sectors and output modalities.
  • Used routing-distribution divergence, dispersion analysis, and layerwise decomposition to construct specialized expert banks that outperformed one-size-fits-all pruning strategies.

Post-Training Interventions

  • Developed and evaluated on-policy distillation and contrastive activation steering as complementary approaches for recovering capabilities lost under aggressive MoE pruning.
  • Used successful teacher rollouts and pruned-model failure trajectories to address repetitive undesirable behavior loops and derive targeted layerwise interventions.