IBM researcher unifying GenAI policy enforcement, LLM safety alignment, and privacy-preserving machine learning
2
papers here
2
reviewed by Evolve
1
first-author
1
corresponding
—
works · OpenAlex
23
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 is not scored in this snapshot: the underlying per-work records needed to compute it are insufficient. This axis would measure total and per-year publication volume, such as the 32 works logged over the last three years.
Impact—
Impact is not scored here: citation-stream data for this axis is insufficient in the current snapshot. It would measure citation volume and influence; only headline counts (h-index 23, i10-index 37) are available, and no score is estimated.
Focus0/100
Focus scores 0, reflecting an even 50/50 field split between AI security and governance and LLM safety alignment rather than narrow concentration. Topic counts reinforce this breadth: privacy-preserving technologies in data lead with 71, ahead of adversarial robustness (29), stochastic gradient optimization (20), ethics and social impacts of AI (18), and cryptography and data security (17) — a portfolio spanning policy, alignment, and underlying ML and security techniques.
Collaboration41/100
Collaboration scores 41, a moderate level. The snapshot records 7 distinct coauthors, with recurring partners including Roberto Campbell, Momin Abbas, Muneeza Azmat, Michal Ulewicz, Raya Horesh, Kristjan Greenewald, and Rogério Abreu de Paula. The two reviewed 2026 papers show both authorship modes — first-and-corresponding on the GenAI policy position paper and last-author seniorship on the aLoRA harm mitigation framework — consistent with an industrial research lab pattern at IBM.
Activity100/100
Activity scores a full 100. OpenAlex logs works in every year from 2011 (2 works) through 2026 except 2015, with output rising from 9 works in 2019 to a peak of 18 in 2022, then 11 in 2024, 11 in 2025, and 10 already logged in 2026 — 32 works in the last three years. Both reviewed papers date from 2026, confirming the record is current, not archival.
Reproducibility20/100
Reproducibility scores 20 (reproMean) across the 2 papers with reproducibility assessments, both rated low readiness. The first-author GenAI policy paper is a position paper judged persuasive but unvalidated, with no empirical validation of its enforcement framework. The aLoRA harm mitigation paper shows solid routing gains, but Evolve notes its alignment wins rely on LLM-made labels and match plain granite-instruct baselines, limiting confidence in the reported improvements.
Research summary
Nathalie Baracaldo is a researcher at IBM (United States) working across AI security and governance and LLM safety alignment, with the field split evenly (0.5/0.5) between the two. An OpenAlex bibliometric snapshot reports an h-index of 23 and an i10-index of 37, built on a record stretching back to 2011 (2 works logged) that accelerates from 9 works in 2019 to a peak of 18 in 2022, with 11 in 2024, 11 in 2025 and 10 already in 2026 — 32 works in the last three years, earning a maximum activity score of 100. Topic counts are led by privacy-preserving technologies in data (71), followed by adversarial robustness in machine learning (29), stochastic gradient optimization (20), ethics and social impacts of AI (18), and cryptography and data security (17). In the Evolve catalog, Baracaldo first-authored the 2026 position paper 'From Alignment to Access Control', arguing that prompt-based GenAI policy enforcement is unsafe and hard access controls are needed — judged persuasive but unvalidated — and co-authored (last author) the aLoRA-based modular harm mitigation framework, whose routing gains are solid although its alignment wins rest on LLM-made labels. Reproducibility readiness across both reviewed papers is rated low (mean 20).
Collaboration pattern
Collaboration is moderate (score 41) with 7 distinct coauthors in the snapshot. The frequent-coauthor list — Roberto Campbell, Momin Abbas, Muneeza Azmat, Michal Ulewicz, Raya Horesh, Kristjan Greenewald, and Rogério Abreu de Paula — suggests a stable IBM research group. The two reviewed 2026 papers bracket both ends of the author hierarchy: first-and-corresponding on the GenAI policy position paper, last-author seniorship on the aLoRA harm mitigation work, consistent with leading some projects and mentoring others.
Research keywords
Highlights
GenAI policy enforcement needs hard access controls, not prompts
Modular harm mitigation via aLoRA router and aligners
Timeline
- 2026First author and corresponding author of 'From Alignment to Access Control: A Framework for GenAI Policy Enforcement' (arXiv), arguing that prompt-based GenAI policy enforcement is unsafe and that hard access controls are needed.
- 2026Last author of 'An Efficient and Modular Framework for Targeted Harm Mitigation in LLMs' (arXiv), introducing an Activated LoRA (aLoRA) router with modular aligners for targeted harm correction.
- 2026OpenAlex logs 10 works for 2026 so far, on top of 11 in 2024 and 11 in 2025 (32 works over the last three years), sustaining a maximum activity score of 100.
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.
- Bibliometric statistics come from an OpenAlex identity snapshot; the output and impact radar axes are unscored because the underlying per-work and citation-stream data is insufficient, and h-index (23) / i10-index (37) are headline counts only.
- Only 2 papers from the Evolve catalog were reviewed, both 2026 arXiv preprints; earlier work is reflected solely in aggregated bibliometrics.
- Both reviewed papers are rated low reproducibility readiness (reproMean 20); verdicts reflect Evolve review judgments, not independent replication.
- Timeline is restricted to years present in the provided paper records; this profile is name-resolved from public bibliographic data (OpenAlex, ORCID 0000-0001-9469-045X).