Peisong Wang

Research profile

Research summary

Peisong Wang is an early-stage researcher working at the intersection of computer systems and machine-learning efficiency. The OpenAlex snapshot shows an even field split — 50% computer systems (computer architecture and graphics), 50% machine learning (efficient LLM inference) — and two 2026 arXiv preprints cataloged by Evolve, both with Wang in a middle author position. The first, 'From Splats to Silicon,' is a workload-centric survey of 3D Gaussian splatting efficiency with original GPU profiling across rasterization and sorting; Evolve calls it rigorous and unusually honest, while noting a narrow empirical base of one GPU and four scenes, and rates its reproducibility readiness high. The second, an empirical study of dynamic expert pruning in fine-grained MoE LLMs, finds that uniformly keeping 2/3 of the chosen experts retains 98.8% of performance, with dynamic pruning paying off only under aggressive budgets; readiness is medium. Quantitative signals are thin and very recent: one indexed work in 2026, recent3yWorks of 1, and h-index and i10-index recorded as 0, so output and impact are not yet scored. What is visible is promising: collaboration breadth of 23 distinct coauthors (score 63) and a leading reproducibility score of 78, averaged over 2 assessed papers. The unifying theme across both papers is computational efficiency — fewer experts, better-characterized GPU workloads, and cheaper inference.

Research keywords

Highlights

• Uniform 2/3 expert retention keeps 98.8% of MoE performance

  An empirical study of dynamic expert pruning in fine-grained MoE LLMs shows that uniformly keeping 2/3 of the chosen experts retains 98.8% of performance, while dynamic pruning pays off only under aggressive budgets, with results compared at equal budgets. Evolve rates it convincing, solid work with medium reproducibility readiness.

• Workload-centric rethinking of 3DGS computational efficiency

  'From Splats to Silicon' surveys the computational efficiency of 3D Gaussian splatting through workload characterization, backed by original GPU profiling that covers rasterization and sorting and points to hardware accelerator opportunities. Evolve calls it rigorous and unusually honest, while noting a narrow empirical base of one GPU and four scenes; reproducibility readiness is high.

• Early reproducibility signal: 78 averaged over two assessed papers

  The strongest quantitative signal in this early-stage profile is reproducibility: a mean score of 78 across 2 assessed papers, with high readiness for the 3DGS survey and medium for the MoE pruning study, alongside 23 distinct coauthors.

Timeline

1. 2026arXiv preprint 'From Splats to Silicon: Rethinking Computational Efficiency of 3DGS' appears — a workload-centric survey with original GPU profiling, later rated high for reproducibility readiness.
2. 2026arXiv preprint 'You Only Need 2/3 of the Chosen Experts' appears — an empirical study reporting that uniform 2/3 expert retention keeps 98.8% of MoE performance, rated medium for readiness.
3. 2026Both works enter the Evolve catalog, forming an early profile with 2 assessed papers, a reproducibility mean of 78, and 23 distinct coauthors.

Papers in the Evolve catalog

• 2026-09-22

  You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs

• 2026-09-05

  From Splats to Silicon: Rethinking Computational Efficiency of 3DGS

Frequent co-authors