Saadia Gabriel

University of California, Los Angeles · United States

Unclaimed · AI-structured profileORCID OpenAlex

UCLA researcher bridging LLM agent reliability and cultural multimodal AI through benchmark-first evaluation

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—

Not yet scored: the snapshot lacks the complete publication records this axis needs. Output would measure the volume and consistency of published work; partial signals exist (38 works in three years, 15 logged in 2026), but a score awaits fuller indexing.

Impact—

Not yet scored: citation-linked impact data is insufficient in this snapshot. Impact would measure citation-based influence such as normalized citation rates. Raw indicators exist — h-index 14, i10-index 15 — but no dimension score has been computed, so none is estimated here.

Focus0/100

Scores 0 on focus, reflecting maximal dispersion rather than inactivity. The field distribution is split exactly 50/50 between machine learning / LLM agent reliability and multimodal machine learning (cultural AI / NLP), and snapshot topics reinforce the breadth: topic modeling (32), ethics and social impacts of AI (18), misinformation and its impacts (18), NLP techniques (16), and hate speech and cyberbullying detection (12) — a deliberately dual-track, benchmark-centric portfolio.

Collaboration64/100

Collaboration scores 64, built on 24 distinct coauthors in the snapshot. Frequent collaborators include Salman Rahman, Yubin Kim, Mihir Parmar, A. Ali Heydari, Genglin Liu, Simon A. Lee, Weizhi Zhang, Arian Hosseini, Ahmed A. Metwally, Yuzhe Yang, Baharan Mirzasoleiman, and Xin Liu. Both Evolve-reviewed 2026 papers place Gabriel in a middle author position without corresponding status, consistent with a contributor role on team-authored benchmark studies.

Activity100/100

Activity is the strongest axis at 100. The counts-by-year series climbs from 2 works in 2023 to 10 in 2024, 13 in 2025, and 15 in 2026 — three consecutive increases — with 38 works over the last three years. Current output sits at the snapshot's historical peak, fully supporting the maximum activity score.

Reproducibility78/100

Reproducibility averages 78 across the 2 Evolve-reviewed papers. NormViz rates high readiness, backed by a well-validated benchmark; the long-horizon agent failure-localization benchmark rates medium, with draft-state artifacts and an LLM-judged error taxonomy flagged as caveats. Both releases are benchmark-first artifacts, which typically aids replication, though the evidence covers only two papers.

Research summary

Saadia Gabriel is a UCLA-based researcher whose public bibliographic record (resolved via OpenAlex) shows a dual-track machine learning program: LLM agent reliability, and multimodal machine learning with a cultural AI and NLP emphasis. The field split is exactly even at 50/50, and snapshot topics reinforce the breadth: topic modeling (32 works), ethics and social impacts of AI (18), misinformation and its impacts (18), NLP techniques (16), and hate speech and cyberbullying detection (12). Recent activity is intense: 38 works in the last three years, rising from 10 in 2024 to 13 in 2025 and 15 in 2026. Snapshot indicators include an h-index of 14, an i10-index of 15, and 24 distinct coauthors, supporting a moderately collaborative style (collaboration score 64). Evolve has human-reviewed two 2026 arXiv papers. A long-horizon agent benchmark shows frontier LLM judges fail to locate the first error in an agent run while a 4B trained verifier succeeds; NormViz shows vision-language models reach at best 25.3% pair accuracy on visual social norms across global cultures. Both are benchmark-first contributions, and reproducibility across the reviewed pair averages 78 — the agent work at medium readiness, NormViz at high.

Collaboration pattern

With 24 distinct coauthors and a collaboration score of 64, Gabriel works in mid-sized, benchmark-focused teams. Recurring collaborators include Salman Rahman, Yubin Kim, Mihir Parmar, Genglin Liu, Weizhi Zhang, and Arian Hosseini, alongside Ahmed A. Metwally, Yuzhe Yang, Xin Liu, and Baharan Mirzasoleiman, reflecting the breadth of the network. Both Evolve-reviewed papers list Gabriel in a middle author position without corresponding status, pointing to specialist contributions rather than sole-lead roles.

Research keywords

LLM agentslong-horizon reliabilityfailure localizationverifiers and process reward modelsbenchmark design and evaluationvision-language modelscultural and social norm understandingmultimodal reasoningAI ethics and social impactmisinformationhate speech and cyberbullying detectiontopic modeling

Highlights

Frontier LLM judges fail to locate the first error in long agent runs
Evolve's human-verified benchmark shows frontier judges cannot pinpoint where a long-horizon agent run first went wrong, while a 4B trained verifier can — with draft-state artifacts and an LLM-judged error taxonomy noted as caveats.
NormViz: VLMs fall short on visual social norms across cultures
A well-validated benchmark and framework show vision-language models struggle to ground reasoning in global cultural norms, with the best model at 25.3% pair accuracy; synthetic-data finetuning helps, though circularity remains untested.
Three straight years of rising output, peaking in 2026
Counts by year rise from 2 works in 2023 to 10 (2024), 13 (2025), and 15 (2026), with 38 works over the last three years — the snapshot's steepest sustained climb, matching the maximum activity score of 100.

Timeline

2017Earliest activity in the snapshot: 1 work logged in 2017.
2021Output builds to 9 works in 2021, then 10 in 2022 — the first sustained high-output phase.
2023A sharp dip to 2 works logged in 2023, the snapshot's lowest recent year.
2024Rebound year: 10 works, matching the prior 2022 peak and starting a three-year climb.
2026Peak output at 15 works (38 over 2024-2026), including two Evolve-reviewed arXiv papers: the long-horizon agent failure-localization benchmark and NormViz.

Papers in the Evolve catalog

Frequent co-authors

Salman Rahman×1 Yubin Kim×1 Mihir Parmar×1 A. Ali Heydari×1 Genglin Liu×1 Simon A. Lee×1 Weizhi Zhang×1 Arian Hosseini×1 Ahmed A. Metwally×1 Yuzhe Yang×1 Baharan Mirzasoleiman×1 Xin Liu×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 snapshot sourced from OpenAlex; identity name-resolved from public bibliographic data (ORCID 0009-0001-9353-951X listed).
Output and impact dimensions are unscored: citation-linked per-paper records are insufficient in the snapshot.
Reproducibility mean (78) covers only the 2 Evolve-reviewed papers, both 2026 arXiv preprints; venue acceptance is not reflected.
Topic counts reflect the snapshot's clustering and may overlap across categories.