Johns Hopkins control theorist bringing constrained optimization and robustness to LLM alignment and safe fine-tuning
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
15
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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 cannot be scored yet: normalized productivity benchmarks for this profile are unavailable in the current snapshot. This axis would measure publication volume and consistency relative to field and career-stage norms.
Impact—
Impact cannot be scored yet: field-normalized citation metrics are absent from this snapshot. This axis would measure citation influence, uptake, and the relative impact of publications across venues and years.
Focus0/100
A perfect focus score matches the topic distribution: adversarial robustness in machine learning leads with 25 topic records, followed by fault detection and control (20), advanced control systems optimization (18), sparse and compressive sensing (16), and distributed multi-agent control (13) — a coherent control-and-robustness core. The Evolve field distribution is even narrower: the cataloged work falls in a single field, LLM alignment and test-time compute.
Collaboration36/100
The low collaboration score reflects a compact network: only 3 distinct coauthors appear in the last three years alongside 28 works. The recurring partners are Taha Entesari, Jingyu Zhang, and Daniel Khashabi, and both cataloged 2026 papers place Fazlyab in the last author position — consistent with a senior supervisory role on small teams rather than large consortium-style authorship.
Activity99/100
Near-maximal activity is grounded in the counts-by-year record: 28 works in the last three years, with sustained output from 2020 onward — 11 works in 2020, 9 in 2021, 7 in 2022, 5 in 2023, 11 in 2024, 9 in 2025, and 8 already in 2026, including two arXiv papers reviewed by Evolve.
Reproducibility55/100
The reproducibility score of 55 is computed from a small base — one paper assessed (reproMean 55) — so it should be read as preliminary. Both Evolve-reviewed 2026 papers carry a medium reproducibility rating; reviews note that same-reward-model evaluation in the steering paper and mixed results on Qwen-9B in the fine-tuning paper leave replication claims to be confirmed by independent runs.
Research summary
Mahyar Fazlyab is a researcher at Johns Hopkins University whose work sits at the intersection of control theory, optimization, and trustworthy machine learning. OpenAlex records show an h-index of 15 and an i10-index of 23, with 28 works in the last three years and sustained output since 2020 (11 works in 2020 and again in 2024, 9 in 2021 and 2025, 8 so far in 2026). His topic profile is tightly concentrated: adversarial robustness in machine learning leads with 25 records, followed by fault detection and control systems (20), advanced control systems optimization (18), sparse and compressive sensing (16), and distributed multi-agent control (13). His most recent direction transfers this control-theoretic toolkit to LLM alignment. Two 2026 arXiv papers — both with Fazlyab as last author — were reviewed by Evolve: 'Minimally Invasive Steering of Language Models' proposes Fisher-regularized pre-logit steering (MISVO) with first-order KL theory, judged promising though its best-so-far claim is weakened by same-reward-model evaluation; 'Beyond Average Safety' applies chance-constrained fine-tuning with a control-barrier-function-style gradient filter to limit harmful fine-tuning, rated strong overall but mixed on Qwen-9B. Collaboration remains small-scale, with three distinct recent coauthors.
Collaboration pattern
Fazlyab's recent collaboration footprint is compact: 3 distinct coauthors over the last three years while producing 28 works. The recurring partners — Taha Entesari, Jingyu Zhang, and Daniel Khashabi — anchor both 2026 Evolve-reviewed papers, where Fazlyab appears as last (senior) author, non-corresponding. The pattern suggests a small, stable Johns Hopkins-centered group applying control-theoretic methods to LLM alignment rather than large multi-institution consortia.
Research keywords
Highlights
Minimally invasive steering of language models (MISVO)
Chance-constrained fine-tuning against harmful fine-tuning
A control-theoretic toolkit for LLM alignment
Both 2026 papers recast alignment as constrained optimization — entropic risk, chance constraints, Fisher-regularized KL — extending the robustness-and-control program visible in the topic profile (adversarial robustness, 25 records; fault detection and control, 20).
Timeline
- 2013First work recorded in the OpenAlex counts-by-year series (1 work), marking the start of the visible publication record.
- 2017Four works recorded this year, matching the earlier 2015 level and continuing a gradual build-up through the mid-2010s.
- 2020Output jumps to 11 works, the first peak in the record and the start of a sustained high-productivity phase.
- 2024A second peak of 11 works, equaling the 2020 high.
- 2026Eight works recorded so far, including two arXiv papers on LLM alignment and safe fine-tuning reviewed by Evolve, both with Fazlyab as last author.
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 counts, h-index, and topic distribution come from OpenAlex only; other databases were not consulted.
- Output and impact dimension scores are missing in this snapshot; only raw h-index (15) and i10-index (23) are available, with no field-normalized citation percentiles.
- The reproducibility score rests on a single assessed paper (reproMean 55 from 1 paper), so it is preliminary.
- Profile is name-resolved from public bibliographic data (OpenAlex; ORCID 0000-0001-9695-6178); career facts beyond these records are not covered.