Johannes Bjerva

Aalborg University · Denmark

Unclaimed · AI-structured profileORCID OpenAlex

Researcher spanning LLM reasoning, AI safety measurement, and privacy-preserving ML at Aalborg University.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

No output score is computed yet. This axis would measure publication volume and productivity trend; the snapshot records per-year counts and 77 works in 2024-2026, but no validated Evolve output metric exists.

Impact—

No impact score is available yet. This axis would measure field-normalized citation influence; the snapshot lists h-index 18 and i10-index 28, but the normalized benchmarks needed to score it are missing.

Focus0/100

The focus score of 0 signals an unusually broad spread. The field distribution splits evenly at 0.5 between natural language processing (LLM reasoning and AI safety measurement) and computer security and privacy (privacy-preserving machine learning). Topic counts reinforce the breadth: topic modeling (98) and NLP techniques (91) lead, with syntax, semantics and linguistic variation (34), categorization, perception and language (23), and adversarial robustness (21) adding further distinct clusters.

Collaboration41/100

The snapshot records 7 distinct co-authors, with repeated collaborators including Qiongxiu Li, Wenrui Yu, Xiaoyu Luo, Tao Ren, Xiao Li, and Changlong Ji. A collaboration score of 41 reflects a relatively compact co-author network rather than a large distributed one. Recent papers pair him with small teams working on LLM reasoning and on privacy-preserving federated learning, suggesting project-driven collaboration clusters.

Activity100/100

The activity score of 100 is grounded in per-year publication counts: 14 works in 2023, 23 in 2024, and 29 in 2025, with 25 recorded so far in 2026 and 77 works across the last three years. This contrasts with a quieter 2021-2022 stretch (2 works each) and echoes an earlier 2017 peak of 14, marking a pronounced current acceleration.

Reproducibility55/100

Reproducibility averages 55 across the 2 Evolve-evaluated papers, and both carry a 'medium' reproducibility readiness rating. The mid-range score fits the evidence: hidden chain-of-thought extraction requires frontier-model tool access, while the lattice-based reconstruction attack works only for the first aggregation round in small sparse graphs under heavy collusion, making artifacts plausible but demanding to replicate.

Research summary

Johannes Bjerva is a researcher at Aalborg University (Denmark; ORCID 0000-0002-9512-0739) whose OpenAlex-resolved profile shows an unusually wide research span. The field distribution splits evenly (0.5 each) between natural language processing — centered on LLM reasoning and AI safety measurement — and computer security and privacy, focused on privacy-preserving machine learning. Topic counts are led by topic modeling (98) and NLP techniques (91), with further clusters in syntax, semantics and linguistic variation (34), categorization, perception and language (23), and adversarial robustness (21). The snapshot records an h-index of 18 and an i10-index of 28, plus a pronounced late surge: 14 publications in 2023, 23 in 2024 and 29 in 2025, with 77 works counted across 2024-2026 and a maximal activity score of 100. Two 2026 arXiv preprints reviewed by Evolve illustrate the dual track: a last-author study eliciting hidden chain-of-thought in frontier models, judged credible though the traces' equivalence to internal reasoning remains unproven, and a middle-author lattice-based reconstruction attack on secure aggregation in decentralized federated learning, effective only in small sparse graphs in the first round under heavy collusion. Collaboration is compact — 7 distinct co-authors, score 41 — and reproducibility averages 55 across the two evaluated papers.

Collaboration pattern

The snapshot records only 7 distinct co-authors (collaboration score 41), a compact network. Frequent collaborators include Qiongxiu Li — appearing twice in the deduplicated list — together with Wenrui Yu, Xiaoyu Luo, Tao Ren, Xiao Li, and Changlong Ji. Recent work splits into two small teams: one on hidden chain-of-thought in frontier models, one on lattice-based privacy attacks in decentralized federated learning, consistent with project-driven rather than large-lab collaboration.

Research keywords

Large language model reasoningChain-of-thought interpretabilityAI safety measurementDecentralized federated learningSecure aggregationGradient inversion attacksLattice-based cryptanalysisHidden subset sum problemPrivacy-preserving machine learningTopic modelingAdversarial robustnessLinguistic variation

Highlights

Hidden chain-of-thought extraction from frontier models
As last author of this 2026 arXiv study, he contributes to tool-forcing elicitation that surfaces near-native hidden reasoning in frontier models. Evolve's verdict finds the study credible while cautioning that the extracted traces' identity with internal chain-of-thought remains unproven; reproducibility readiness is rated medium.
Lattice-based reconstruction attacks on secure aggregation
In this 2026 arXiv paper (middle author), secure-aggregation leakage in decentralized federated learning is reduced to lattice hidden subset sum and hidden lattice problems (HSSP/HLCP). The attack recovers updates in small sparse graphs, but only in the first round and under heavy collusion; reproducibility readiness is medium.
One audit mindset, two targets
The two reviewed papers share a common audit orientation: eliciting what frontier models hide internally, and recovering what secure aggregation hides cryptographically. This dual track mirrors the even 0.5/0.5 field split in the bibliometric record and underlies the focus score of 0.

Timeline

2011-2015Early publication record appears in the bibliometric snapshot with small annual counts: 2 works in 2011, 4 in 2014 and 1 in 2015.
2016-2017Output rises sharply to 9 works in 2016 and an early peak of 14 in 2017, followed by 10 in 2018 and 8 in 2019.
2021-2022A quiet stretch with only 2 works recorded in each year, the lowest annual counts since 2013.
2023-2025Sharp acceleration: 14 works in 2023, 23 in 2024 and 29 in 2025 — the highest annual counts on record — with the recent three-year total reaching 77 works across 2024-2026.
2026Two Evolve-reviewed arXiv preprints appear — hidden chain-of-thought extraction in frontier models (last author) and lattice-based reconstruction attacks on decentralized federated learning (middle author) — with 25 works recorded year-to-date.

Papers in the Evolve catalog

Frequent co-authors

Wenrui Yu×2 Xiaoyu Luo×1 Tao Ren×1 Xiao Li×1 Qiongxiu Li×1 Changlong Ji×1 Qiongxiu Li×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric metrics (h-index, i10-index, per-year counts, topics, co-authors) come from an OpenAlex identity snapshot and may reflect record-merging artifacts.
Output and impact radar scores are unavailable (dimMissing); h-index 18 and i10-index 28 are raw snapshot values, not field-normalized impact measures.
Qualitative judgments rest on only two Evolve-reviewed papers (both 2026 arXiv preprints); earlier career history, awards and appointments are outside the data scope.
Counts for 2026 are year-to-date and may not represent a complete calendar year.