Johns Hopkins researcher applying constrained optimization and control theory to LLM safety and test-time alignment
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
2
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—
Publication-output data is insufficient in this snapshot, so no score is given. This axis would measure total works, per-year output, and venue distribution derived from a complete, disambiguated publication record.
Impact—
Impact cannot be scored yet: the citation record is too thin (h-index 2, i10-index 0) for a stable estimate. This axis would measure citations, citing works, and downstream influence of the research.
Focus0/100
The focus axis currently scores 0 on the 0-100 scale. The raw field distribution splits evenly — 0.5 for machine learning (LLM safety and constrained optimization) and 0.5 for machine learning (LLM alignment and test-time compute) — two closely adjacent areas. Topic data shows concentration on adversarial robustness (10 assignments) ahead of formal methods in verification (4), stochastic gradient optimization (4), control systems optimization (3), and data classification (3).
Collaboration36/100
The collaboration axis scores 36. Raw indicators show 5 distinct coauthors over the last three years across 7 recent works, with recurring partners Mahyar Fazlyab, Jingyu Zhang, and Daniel Khashabi. Both Evolve-catalog papers are first-authored without a corresponding-author role, consistent with a small, stable team pattern rather than a broad external network.
Activity61/100
The activity axis scores 61. Per-year counts from the snapshot are 2020: 1, 2022: 2, 2023: 3, 2024: 3, 2025: 1, and 2026: 3 works, with 7 works in the last three years. The record shows sustained recent output after a quieter 2025, including two first-author arXiv preprints in 2026.
Reproducibility55/100
The reproducibility axis scores 55, computed from 2 reviewed papers with a mean readiness of 55. Both Evolve-catalog papers are rated medium on reproducibility readiness. Evolve's verdicts note that reported gains partly rest on best-so-far metrics and same-reward-model evaluation in one paper, and on mixed results for Qwen-9B in the other, both tempering replication confidence.
Research summary
Taha Entesari is a Johns Hopkins University-affiliated researcher working at the interface of large language model (LLM) safety and constrained optimization, with a toolkit drawn from control theory, stochastic optimization, and formal verification. This profile is name-resolved from OpenAlex bibliographic data, which attributes an h-index of 2 and an i10-index of 0 — a very early citation record — alongside 7 works in the last three years and 5 distinct coauthors. The topical profile is led by adversarial robustness in machine learning (10 topic assignments), followed by formal methods in verification (4) and stochastic gradient optimization (4), with advanced control systems optimization (3) and machine learning data classification (3). Evolve's field distribution splits evenly (0.5/0.5) between LLM safety with constrained optimization and LLM alignment with test-time compute. Two 2026 first-author arXiv preprints anchor the catalog. 'Beyond Average Safety' frames safe fine-tuning as chance-constrained optimization with a control-barrier-function-style gradient filter to limit safety regressions under harmful fine-tuning; 'Minimally Invasive Steering' proposes Fisher-regularized pre-logit steering (MISVO) with first-order KL theory for test-time alignment. Evolve rates both papers medium on reproducibility readiness, judging the fine-tuning work strong overall (mixed on Qwen-9B) and the steering work promising but with weakened claims from best-so-far metrics and same-RM evaluation.
Collaboration pattern
Entesari's collaboration footprint is compact: 5 distinct coauthors across 7 works in the last three years, with Mahyar Fazlyab, Jingyu Zhang, and Daniel Khashabi as recurring partners. Both Evolve-catalog papers are first-authored without corresponding-author credit, pointing to a small, stable team pattern. The collaboration dimension scores 36 on the 0-100 radar.
Research keywords
Highlights
Chance-constrained fine-tuning to curb harmful fine-tuning
Minimally invasive test-time steering (MISVO)
Timeline
- 2020First publication year recorded in the OpenAlex snapshot, with 1 work.
- 20222 works recorded in this year.
- 20233 works recorded, stepping up from 2022.
- 20243 works recorded, matching the previous year's pace.
- 20251 work recorded — the quietest year since 2020 in the snapshot.
- 20263 works, including both first-author arXiv preprints in the Evolve catalog: MISVO steering and chance-constrained safe fine-tuning.
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-sourced snapshot; no ORCID is linked, so name disambiguation may be imperfect.
- Citation-impact signals are thin (h-index 2, i10-index 0); the output and impact radar axes are left unscored.
- Only 2 papers carry Evolve review verdicts; the wider publication record has not been reviewed.
- Per-year and topic counts reflect only works indexed in this snapshot.