NLP and AI-agent-safety researcher: calibrated small-model scam-call screening and measurement of agentic PDF forgery
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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: beyond the two Evolve-reviewed papers, no publication counts or venue metrics are available. This axis would measure overall research output, productivity over time, and publication venues once a fuller bibliographic record is linked.
Impact—
Not yet scored: the input contains no citation counts, h-index, or adoption signals. This axis would measure how widely the two reviewed preprints are cited and built upon, once citation data accumulates.
Focus0/100
Scores 0 on this axis, a direct read-out of the raw field distribution: the reviewed output splits exactly 50/50 between "AI agent safety and document forensics" and "natural language processing (NLP)". The two 2026 papers tackle different problems — calibrated scam-call screening with small language models versus measuring agentic PDF forgery — so the profile currently reads as evenly balanced across two areas rather than concentrated; a larger sample would show whether this is genuine breadth or early-stage exploration.
Collaboration48/100
Scores 48, computed from 10 distinct coauthors across the reviewed record. Recurring names — Simiao Ren, Ankit Raj, Tommy Duong, Yuxin Zhang, Dennis Ng, Xingyu Shen, Yuchen Zhou, and Neo Tiangratanakul — with Ankit Raj and Xingyu Shen each appearing twice in the coauthor list, point to a small repeat-collaboration core. On both catalog papers Zewde sits in a middle authorship position and is not the corresponding author, consistent with a contributor rather than lead role.
Activity—
Not yet scored: the record covers only papers dated 2026, with no multi-year publication history to establish a cadence. This axis would measure publication frequency and continuity over time, requiring a longer publication window to compute.
Reproducibility55/100
Scores 55, the mean repro-readiness across the two reviewed papers (reproPapers = 2, both rated medium). The forgery paper documents a rule-based scoring pipeline and reports success rates at two thresholds (81.1% raw, 46.2% strict), while the screening paper is candid about test-set exposure but relies entirely on synthetic data. This mixed picture — transparent evaluation design alongside synthetic-only or server-dependent setups — yields a moderate rather than high reproducibility signal.
Research summary
Kidus Zewde is an early-stage researcher whose Evolve profile rests on two 2026 arXiv preprints at the intersection of natural language processing and AI agent safety. The first, "Open-Jev Judgments on CallScreenBench," fine-tunes a JevLite typed-decision readout onto Qwen3-4B for calibrated, one-pass scam-call screening; Evolve's review finds it non-inferior to an LLM judge at roughly 30x lower latency, while noting that all evaluation data is synthetic. The second, "Agents That Edit Documents," measures how reliably LLM agents forge archived PDFs against a non-agentic control, reporting an 81.1% raw success rate that drops to 46.2% under strict filtering; its rule-based scoring design is credited, with server-side heterogeneity flagged as a ranking confound. On both papers Zewde holds a middle authorship position and is not the corresponding author. The bibliometric signal is correspondingly thin: radar dimensions currently score collaboration at 48 (10 distinct coauthors, including repeat collaborators such as Ankit Raj and Xingyu Shen), focus at 0 (a perfect 50/50 split between "AI agent safety and document forensics" and NLP), and reproducibility at 55 (a medium rating on both reviewed papers). Output, impact, and activity remain unscored because the underlying data is insufficient. Institution and country are not recorded, so this profile is name-resolved from public bibliographic data under ORCID 0009-0000-0113-7963.
Collaboration pattern
Ten distinct coauthors appear across the record: Simiao Ren, Ankit Raj, Tommy Duong, Yuxin Zhang, Dennis Ng, Xingyu Shen, Yuchen Zhou, and Neo Tiangratanakul recur, with Ankit Raj and Xingyu Shen each listed twice, hinting at a small repeat-collaboration core. Zewde holds middle, non-corresponding authorship on both papers, consistent with contributing within a team rather than leading it; the collaboration radar score of 48 sits in the middle range.
Research keywords
Highlights
Small-model scam screening at ~30x lower latency
Quantifying how reliably LLM agents forge archived PDFs
Timeline
- 2026Co-authored "Open-Jev Judgments on CallScreenBench", an arXiv preprint on calibrated one-pass scam-call screening with a small language model (middle authorship, not corresponding).
- 2026Co-authored "Agents That Edit Documents", an arXiv benchmark measuring agentic PDF forgery against a non-agentic control (middle authorship, not corresponding).
- 2026Both catalog papers appeared in the same year, splitting the reviewed output evenly between NLP screening and agent-safety document forensics.
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.
- Profile is name-resolved from public bibliographic data under ORCID 0009-0000-0113-7963; institution and country are not recorded in the available inputs.
- All quantitative signals derive from only two Evolve-reviewed 2026 arXiv preprints; citation counts, h-index, and venue history are outside the provided data.
- Output, impact, and activity radar dimensions are unscored (dimMissing) because the underlying bibliometric data is insufficient; no estimates are substituted.
- Reproducibility reflects a mean of 55 over just two papers, both rated medium readiness; the 50/50 field split and focus score rest on the same small sample.