Wenxuan Zeng

Research profile

Research summary

Wenxuan Zeng is a researcher at Sun Yat-sen University (China) whose work splits evenly between two fast-moving areas: privacy-preserving machine learning — combining applied cryptography with computer architecture — and machine learning for LLM training data and scientific code generation. The Evolve catalog includes two 2026 arXiv papers with Zeng in a middle authorship position. OptiPrime applies protocol–hardware co-design to private inference, cutting latency up to 5.7x versus Cheetah through an FPGA implementation of BFV-based homomorphic encryption. SciWalker synthesizes 8,178 scientific coding problems using operator graphs and execution feedback, then trains with reinforcement learning (GSPO) for gains on scientific code benchmarks. OpenAlex topics reinforce this dual track: cryptography and data security (9 works), privacy-preserving technologies in data (7), and topic modeling (6). Activity is the standout signal — a score of 83, with output rising from 1 work in 2022 to 8 in 2026 and 16 works in the last three years, alongside 14 distinct coauthors. Reproducibility averages 55 across the two reviewed papers: strong hardware evaluation in one case, but closed platforms and missing ablations or error bars cap readiness. Citation impact is not yet scoreable, with an h-index of 5 and i10-index of 4 in the snapshot. This is an early-stage but rapidly scaling profile.

Research keywords

Highlights

• SciWalker: 8,178 synthetic scientific coding problems with strong RL gains

  The paper synthesizes scientific coding problems via operator graphs and execution feedback — 8,178 problems in total — and trains with reinforcement learning (GSPO), reporting strong gains. Evolve notes the missing ablations and error bars and that the checkpoint was picked on the SciCode test set. Reproducibility readiness: medium.

• OptiPrime: up to 5.7x faster private inference via protocol–FPGA co-design

  Protocol–hardware co-design cuts private-inference latency up to 5.7x versus Cheetah, built on homomorphic encryption (BFV) and secure multi-party computation with an FPGA implementation. Evolve credits the real-hardware evaluation but flags a closed platform that limits reproducibility. Reproducibility readiness: medium.

• An even split between secure computation systems and LLM-for-science

  The OpenAlex field distribution assigns 50% to privacy-preserving machine learning and 50% to LLM training data and scientific code generation, with topic counts led by cryptography and data security (9), privacy-preserving technologies (7), and topic modeling (6) — two parallel research lines clearly visible across the two reviewed papers.

Timeline

1. 2022First recorded output year in the OpenAlex snapshot, with 1 work.
2. 20233 works recorded — the early accumulation phase of the profile.
3. 2024Output rises to 5 works, the strongest pre-2026 year in the snapshot.
4. 20253 works recorded in the snapshot.
5. 2026Peak year with 8 works; both reviewed arXiv papers — SciWalker and OptiPrime — appear.

Papers in the Evolve catalog

• 2026-09-24

  SciWalker: Synthesizing Scientific Coding Problems with Operator Graphs and Execution Feedback

• 2026-09-15

  OptiPrime: Optimizing Private Inference through Protocol-Hardware Co-design

Frequent co-authors