Wenrui Yu

Research profile

Research summary

Wenrui Yu is an early-career researcher at Aalborg University, Denmark (ORCID 0009-0008-4089-3727), whose profile is name-resolved from OpenAlex bibliographic data. The snapshot shows an even 50/50 split between natural language processing (LLM reasoning and AI safety measurement) and computer security and privacy (privacy-preserving machine learning). Topic counts are led by Cryptography and Data Security and Privacy-Preserving Technologies in Data (10 each), followed by Distributed Control Multi-Agent Systems (5), Internet Traffic Analysis and Secure E-voting (3), and Stochastic Gradient Optimization Techniques (2). Output is small but accelerating: 12 works in the recent three-year window, with counts by year of 3 (2024), 3 (2025) and 6 (2026); reported h-index is 2 with i10-index 1. Evolve has reviewed two 2026 arXiv papers. As first author, 'When Topology Betrays Privacy' reduces secure-aggregation leakage in decentralized federated learning to lattice hidden subset sum problems via a Nguyen–Stern-style attack; Evolve judges the study credible but notes recovery is limited to small sparse graphs, the first round, and heavy collusion. As a middle author, 'Capable yet Parsimonious' extracts hidden chain-of-thought from frontier models via tool-forcing, with trace-to-internal-CoT identity left unproven. Collaboration rests on 7 distinct coauthors; the reproducibility mean across the two assessed papers is 55, with both rated medium readiness.

Research keywords

Highlights

• First-author lattice attack on secure aggregation in decentralized federated learning

  The 2026 arXiv paper reduces secure-aggregation leakage in decentralized federated learning to lattice HSSP/HLCP and applies a Nguyen–Stern-style reconstruction. Evolve finds the study credible but notes the attack recovers updates only in small sparse graphs, in the first round, and under heavy collusion. ReproReadiness: medium.

• Middle-author study extracting hidden chain-of-thought from frontier models

  In this 2026 arXiv collaboration, tool-forcing elicits near-native hidden reasoning in frontier models. Evolve rates the study credible while flagging that the identity of extracted traces with internal chain-of-thought remains unproven. ReproReadiness: medium.

• Output accelerating: 12 works in three years, 2026 already at 6

  OpenAlex counts-by-year rise from 3 (2024) and 3 (2025) to 6 (2026), with the recent three-year total at 12 works. Reported h-index is 2 with i10-index 1 — an early-career signal that is still forming.

Timeline

1. 2024OpenAlex counts-by-year records 3 works — the earliest year in the snapshot.
2. 20253 works recorded, matching the previous year's pace.
3. 2026Counts-by-year reaches 6 works, double either earlier year; the recent three-year total stands at 12.
4. 2026First-author arXiv paper 'When Topology Betrays Privacy' appears; Evolve judges it credible with attack scope bounded to small sparse graphs, first round and heavy collusion.
5. 2026Middle-author arXiv paper 'Capable yet Parsimonious' appears; Evolve notes the trace-to-internal-CoT identity remains unproven.

Papers in the Evolve catalog

• 2026-09-22

  Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models

• 2026-09-08first author

  When Topology Betrays Privacy: Lattice-Based Reconstruction Attacks on Secure Aggregation in Decentralized Federated Learning

Frequent co-authors