Wojciech Samek
Fraunhofer Institute for Telecommunications, Heinrich Hertz Institute · Germany
AI safety & XAI researcher at Fraunhofer HHI: LLM agent verification, training data attribution, robustness
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Output volume is not scored yet: this OpenAlex record lacks a complete, deduplicated publication list for the name, so total works cannot be reliably computed. This axis would measure cumulative publication count and per-year productivity.
Impact—
Impact is not scored yet: citation data here is unreliable — the record shows an h-index of 1 and i10-index of 0 despite 33 recent works, indicating under-attribution. This axis would measure citations, h-index, and scholarly influence.
Focus0/100
The focus score is 0, reflecting a broad portfolio rather than a narrow specialty: the field distribution splits 50/50 between LLM-agent interpretability/safety and ML work on LLM safety and training data attribution, while topics span XAI (17), adversarial robustness (10), privacy-preserving technologies (9), topic modeling (6), and software testing and debugging (5).
Collaboration62/100
Collaboration scores 62, drawn from 22 distinct coauthors across 33 works in the last three years — a moderately collaborative, team-based pattern. Twelve recurring partners appear, including Tobias Labarta, Frederik Pahde, Maximilian Dreyer, Sebastian Lapuschkin, and David Birkenberger, suggesting stable in-house teams plus external academic collaborators.
Activity100/100
Activity scores 100: counts-by-year jump from 1 (2018) and 1 (2022) to 4 (2025) and 29 (2026), with 33 works in the last three years — an accelerating publishing rhythm concentrated almost entirely in the most recent year.
Reproducibility78/100
Reproducibility scores 78, the mean over 2 cataloged papers with Evolve reproducibility assessments. The NetOps-agent verification paper is rated high readiness (symbolic replay yields verified per-action ground truth), while the subliminal-learning attribution study is rated medium, reflecting its narrow animal-preference setup.
Research summary
Wojciech Samek is a researcher profile name-resolved from OpenAlex bibliographic data, affiliated with the Fraunhofer Institute for Telecommunications, Heinrich Hertz Institute (Germany). The record shows highly active recent output: 33 works in the last three years, with counts-by-year rising from 4 works in 2025 to 29 in 2026. Topical clustering concentrates on explainable artificial intelligence (17 tagged works), adversarial robustness in machine learning (10), and privacy-preserving technologies (9), with secondary activity in topic modeling (6) and software testing and debugging (5). The field distribution splits evenly (50/50) between LLM-agent interpretability and safety, and machine learning work on LLM safety and training data attribution. Three 2026 arXiv preprints appear in the Evolve catalog: a critique of whether gradient-based data attribution can filter subliminal learning, an AttnLRP-based framework for verifying NetOps agents at action-level granularity, and Modelpedia, an LLM-built catalog of over 1,000 model findings from ICLR 2024-25. Two of these carry Evolve reproducibility assessments averaging 78/100. Recorded citation indicators (h-index 1, i10-index 0) are inconsistent with this output volume and are treated as incomplete; the collaboration score of 62 reflects 22 distinct coauthors.
Collaboration pattern
The pattern is team-based and moderately broad: 22 distinct coauthors across 33 works in the last three years (collaboration score 62), with 12 recurring partners including Tobias Labarta, Frederik Pahde, Novak Boškov, Maximilian Dreyer, David Birkenberger, Sebastian Lapuschkin, Manzoor Ahmed Khan, and Luca Longo. This mix suggests a stable in-house group at Fraunhofer HHI plus recurring external academic collaborators, rather than one-off tie-ups.
Research keywords
Highlights
Data attribution filters subliminal learning only unreliably
Action-level safety signals for verifying NetOps agents
Modelpedia catalogs 1,000+ model findings for AI meta-science
Timeline
- 2018Earliest recorded work in this OpenAlex counts-by-year profile (1 work); the record is sparse in these early years.
- 20221 recorded work; output remains sparse before 2025 in this name-resolved record.
- 2025Output picks up: 4 recorded works, marking the start of the current active phase.
- 2026Peak activity: 29 recorded works, including three Evolve-reviewed arXiv preprints on data attribution versus subliminal learning, action-level NetOps agent verification, and the Modelpedia meta-science catalog.
Papers in the Evolve catalog
Frequent co-authors
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.
- All bibliometrics come from an OpenAlex name-resolved record (no ORCID provided); the recorded h-index of 1 and i10-index of 0 are inconsistent with 33 recent works and likely reflect incomplete author disambiguation.
- Output and impact radar axes are not scored because citation and complete publication-list data are insufficient in the current record.
- The three reviewed papers are 2026 arXiv preprints; peer-review status and final venues are not captured in the provided data.
- Frequent-coauthor relationships are derived from the OpenAlex record; paper-level coauthor lists were not provided, so coauthors are not mapped to specific papers.