Zihan Wang

Research profile

Research summary

Zihan Wang is an early-career researcher at the University of Science and Technology of China whose public record — name-resolved through OpenAlex, with no linked ORCID — currently comprises two 2026 arXiv preprints reviewed by Evolve. The field distribution splits exactly 50/50 between computer systems (ML systems/operating systems) and speech processing/automatic speech recognition. As first author of SeqMoE, Wang proposes predictive, graph-compatible MoE offloading: sequence-model prediction of expert activations, matroid-based prefetch scheduling, and a CUDA-graph-compatible runtime targeting near-full-load LLM inference; Evolve rates the results strong but the reproducibility readiness low, citing missing variance data, code, and key baselines. As a middle author on AURA, Wang contributes to uncertainty-routed activation editing that reportedly cuts Whisper non-speech hallucination from 89.2% to 1.9% with roughly 500x fewer parameters than LoRA, earning high reproducibility readiness. The radar scores reflect this early, two-track state: activity 32 (2 works, all in 2026), collaboration 55 (15 distinct coauthors), reproducibility 60 (mean over 2 papers), and focus 0, given the even field split. Output and impact dimensions lack sufficient data. The recurring thread across both papers is efficient adaptation and deployment of large foundation models — minimizing parameters edited in speech models and memory traffic in MoE serving.

Research keywords

Highlights

• AURA cuts Whisper non-speech hallucination from 89.2% to 1.9%

  Co-authored (middle position) 2026 arXiv work applying uncertainty-routed activation editing to cross-attention heads to acoustically ground speech foundation models, achieving the reported reduction with roughly 500x fewer parameters than LoRA; Evolve calls it promising but Whisper-only, with high reproducibility readiness.

• First-authored SeqMoE targets near-full-load MoE offloading

  2026 arXiv systems paper proposing sequence-model prediction of expert activations, matroid-based prefetch scheduling, and a CUDA-graph-compatible runtime to keep MoE LLM inference near full-load performance; Evolve's review finds the results strong but notes missing variance data, code release, and key baselines.

• A shared thread: efficient adaptation and serving of foundation models

  Across both reviewed papers the underlying problem is the cost of running large models — AURA edits far fewer parameters than LoRA to fix speech hallucination, while SeqMoE reduces memory traffic in MoE inference — which explains the even 50/50 split between speech processing and ML systems in the field distribution.

Timeline

1. 2026Posted first-author preprint SeqMoE on arXiv, proposing predictive, graph-compatible MoE offloading for LLM inference.
2. 2026Co-authored (middle position) the AURA preprint on uncertainty-routed activation editing for speech foundation models.
3. 2026OpenAlex indexes both works in this single year (countsByYear 2026: 2), with recent3yWorks at 2 and an activity score of 32.
4. 2026Evolve reviews both papers, rating AURA's reproducibility readiness high and SeqMoE's low; reproducibility mean is 60 across the 2 papers.

Papers in the Evolve catalog

• 2026-09-21

  AURA: Uncertainty-Routed Activation Editing for Acoustic Grounding in Speech Foundation Models

• 2026-09-11first author

  SeqMoE: Toward Full-Load Performance via Predictive and Graph-Compatible MoE Offloading

Frequent co-authors