Tommy Duong

—

Unclaimed · AI-structured profileOpenAlex

Early-stage researcher in NLP and AI agent safety: calibrated scam-call screening and agentic PDF forgery measurement.

papers here

reviewed by Evolve

first-author

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—

Publication output data is not available in this profile yet. This axis would measure total papers, venue distribution, and output trajectory over time; no volume statistics are currently on record.

Impact—

Citation and impact data is not available in this profile yet. This axis would measure citation counts, scholarly influence, and uptake of the reviewed work; nothing is scored or estimated here.

Focus0/100

Focus scores 0 because the observed field distribution splits exactly evenly — 0.5 natural language processing and 0.5 AI agent safety and document forensics — across the reviewed papers. One line of work builds calibrated screening systems with small language models; the other builds capability benchmarks for agentic document forgery. The two share an evaluation-first methodology but no single dominant specialization, so this axis measures concentration of output across fields and currently reads as fully dispersed.

Collaboration48/100

Collaboration scores 48 based on 10 distinct coauthors across the reviewed record. Frequent collaborators include Simiao Ren, Ankit Raj, Yuxin Zhang, Dennis Ng, Xingyu Shen, Kidus Zewde, Yuchen Zhou, and Neo Tiangratanakul, with Xingyu Shen and Ankit Raj recurring in the frequency list. On both reviewed 2026 arXiv papers the author sits in a middle authorship position and is not the corresponding author, suggesting consistent participation in group projects rather than leading them.

Activity—

Publication activity data is not available in this profile yet. This axis would measure recent publication cadence and year-over-year consistency; only the two 2026 arXiv papers are currently on record.

Reproducibility55/100

Reproducibility averages 55 across the two papers with reproducibility assessments, both rated medium readiness. The scam-screening paper is candid about test-set exposure, which supports replication, but its entire dataset is synthetic, limiting transfer to real call traffic. The PDF-forgery study's rule-based verification is judged rigorous, though server-side heterogeneity across providers interferes with ranking stability. Together these pull the mean to 55 — adequate, but dependent on external model access.

Research summary

Tommy Duong is an early-stage researcher whose reviewed work splits evenly between natural language processing and AI agent safety with document forensics — each field carries a 0.5 share of the observed distribution. Two 2026 arXiv papers are reviewed in the Evolve catalog, both with the author in a middle position and neither as corresponding author. The first applies a JevLite typed-decision readout, LoRA fine-tuned on Qwen3-4B, to CallScreenBench, delivering calibrated one-pass scam-call screening that is non-inferior to an LLM judge at roughly 30x lower latency; Evolve credits its candid treatment of test-set exposure while noting that all data is synthetic. The second measures how reliably LLM agents forge archived PDFs against a non-agentic control, reporting an 81.1% raw forgery success rate that falls to 46.2% under strict filtering; its rule-based scoring is judged rigorous, though server-side heterogeneity muddies ranking. The common thread is a measurement-first style: calibrated or rule-based readouts instead of free-form LLM judgment. Collaboration data records 10 distinct coauthors (score 48), including Simiao Ren, Ankit Raj, Yuxin Zhang, Dennis Ng, Xingyu Shen, Kidus Zewde, Yuchen Zhou, and Neo Tiangratanakul. Reproducibility averages 55 across two papers, both medium readiness. Output, impact, and activity metrics are not yet available.

Collaboration pattern

Tommy Duong collaborates in mid-sized teams, with 10 distinct coauthors recorded (collaboration score 48). The most frequent collaborators are Simiao Ren, Ankit Raj, Yuxin Zhang, Dennis Ng, Xingyu Shen, Kidus Zewde, Yuchen Zhou, and Neo Tiangratanakul, with Xingyu Shen and Ankit Raj appearing repeatedly in the frequency list. On both reviewed 2026 arXiv papers the author holds a middle position and is not corresponding, consistent with contributing inside a rotating group rather than leading a fixed lab.

Research keywords

scam-call detectioncalibrationtyped decisionslabel-token readoutLoRA fine-tuningsmall language modelsagentic AI benchmarkPDF forgerydocument fraudrule-based verificationLLM agentscapability evaluation

Highlights

Calibrated scam-call screening at ~30x lower latency
On CallScreenBench, a JevLite typed-decision readout LoRA fine-tuned on Qwen3-4B produces calibrated one-pass screening judgments that are non-inferior to an LLM judge while cutting latency by roughly 30x; Evolve credits the candid treatment of test-set exposure but notes that all data is synthetic.
Quantifying agentic PDF forgery against a non-agentic control
The study measures how reliably LLM agents forge archived PDFs, reporting an 81.1% raw success rate that drops to 46.2% under strict filtering. Evolve finds the rule-based scoring design rigorous, while flagging that server-side heterogeneity interferes with ranking.
A measurement-first, rule-based evaluation style
Across both papers the shared thread is replacing free-form LLM judgment with typed, calibrated, or rule-based readouts — one for scam screening decisions, one for verifying document forgery — paired with explicit honesty about benchmark limitations such as synthetic data and server-side heterogeneity.

Timeline

2026Co-authored (middle position) the arXiv preprint 'Open-Jev Judgments on CallScreenBench' on calibrated one-pass scam-call screening with a small language model.
2026Co-authored (middle position) the arXiv preprint 'Agents That Edit Documents' measuring agentic PDF forgery against a non-agentic control.
2026Both papers entered the Evolve review catalog, each rated medium reproducibility readiness and averaging 55 on the reproducibility axis.

Papers in the Evolve catalog

Frequent co-authors

Simiao Ren×2 Ankit Raj×2 Yuxin Zhang×2 Dennis Ng×2 Xingyu Shen×2 Kidus Zewde×2 Yuchen Zhou×2 Neo Tiangratanakul×2 Xingyu Shen×1 Ankit Raj×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.

Institution, country, and ORCID are absent from the input; the profile is name-resolved from public bibliographic data and the name may be ambiguous.
Output, impact, and activity scores are missing (dimMissing): no publication counts, citation metrics, or yearly cadence data were available.
Only two Evolve-reviewed papers, both 2026 arXiv preprints with middle-position authorship, underlie this profile; the full publication history is not represented.
Radar scores (collaboration 48, focus 0, reproducibility 55) come from the available bibliometric snapshot; reproducibility is based on 2 papers only.