Neo Tiangratanakul

—

Unclaimed · AI-structured profileOpenAlex

NLP and AI agent safety researcher: small-model scam-call screening and benchmarks for agentic PDF forgery.

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—

Output data is not yet available: this axis would measure publication volume, authorship positions, and productivity over time. Evolve currently catalogues only two 2026 arXiv preprints for this researcher, which is too thin a base to score.

Impact—

Impact data is not yet available: this axis would track citations, downloads, and uptake of the catalogued work. Both papers are recent 2026 arXiv preprints, so citation-based signals have not had time to accumulate.

Focus0/100

Focus scores 0/100, the lowest concentration band, matching the raw field distribution: output splits evenly at 50% “AI agent safety and document forensics” and 50% “natural language processing (NLP)”. The two catalogued 2026 papers sit in these distinct areas — small-model calibration for scam-call screening versus agentic PDF-forgery benchmarks — indicating a deliberately two-sided agenda rather than a single specialisation.

Collaboration48/100

Collaboration scores 48/100, a mid-range value, with a raw count of 10 distinct coauthors across the catalogued papers. The frequent-coauthor list (10 entries, 8 distinct names) covers Simiao Ren, Ankit Raj, Tommy Duong, Yuxin Zhang, Dennis Ng, Xingyu Shen, Kidus Zewde, and Yuchen Zhou, with Ankit Raj and Xingyu Shen each appearing twice — the closest recurring ties. This researcher holds the last author position on both papers.

Activity—

Activity data is missing: this axis would measure publication cadence and year-over-year output. With both catalogued papers dated 2026, there is insufficient longitudinal information to score recent research tempo honestly.

Reproducibility55/100

Reproducibility scores 55/100, computed as a mean over two scored papers (reproMean 55, reproPapers 2), both rated medium readiness. The scam-screening paper is candid about test-set exposure but relies entirely on synthetic data; the PDF-forgery benchmark uses rigorous rule-based scoring, though server-side heterogeneity confounds ranking. Together they suggest careful evaluation hygiene with disclosed limitations.

Research summary

Neo Tiangratanakul is an early-stage profile on Evolve, built from two 2026 arXiv preprints on which this researcher appears as last author (not flagged as corresponding). The catalogued work spans two equally weighted areas — the raw field distribution splits 50/50 between “AI agent safety and document forensics” and “natural language processing (NLP)” — producing a focus score of 0/100, the least concentrated band. The first paper, “Open-Jev Judgments on CallScreenBench”, shows that a JevLite typed-decision readout fine-tuned (LoRA) into Qwen3-4B is non-inferior to an LLM judge for one-pass scam-call screening at roughly 30x lower latency; Evolve notes the authors are candid about test-set exposure but that all evaluation data is synthetic. The second, “Agents That Edit Documents”, quantifies how reliably LLM agents can forge archived PDFs against a non-agentic control: an 81.1% raw success rate, dropping to 46.2% under strict filtering; Evolve praises the rule-based scoring design but flags server-side heterogeneity as a confound for rankings. Bibliometrically, collaboration scores 48/100 with 10 distinct coauthors, and reproducibility scores 55/100 (reproMean 55 over 2 scored papers, both rated medium readiness). Output, impact, and activity axes lack sufficient data and remain unscored. Institution, country, and ORCID are unavailable, so this profile is name-resolved from public bibliographic data.

Collaboration pattern

The catalogued work is collaborative: 10 distinct coauthors appear across the two 2026 papers, yielding a mid-range collaboration score of 48/100. Recurring partners include Simiao Ren, Ankit Raj, Tommy Duong, Yuxin Zhang, Dennis Ng, Xingyu Shen, Kidus Zewde, and Yuchen Zhou; Ankit Raj and Xingyu Shen each appear twice, marking the strongest ties. Both papers place this researcher in the last author position, consistent with a senior supervisory role, though neither flags corresponding authorship.

Research keywords

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

Highlights

Small-model scam screening at ~30x lower latency
On CallScreenBench, a JevLite typed-decision readout fine-tuned into Qwen3-4B is non-inferior to an LLM judge for one-pass scam-call screening at roughly 30x lower latency. Evolve credits the paper's candid treatment of test-set exposure but notes that all evaluation data is synthetic.
First quantitative read on agentic PDF forgery
The benchmark measures how reliably LLM agents can forge archived PDFs against a non-agentic control: an 81.1% raw success rate falling to 46.2% under strict filtering. Evolve finds the rule-based verification design rigorous but flags server-side heterogeneity as a confound for rankings.
A two-sided 2026 agenda pairing rigor with candor
Both catalogued papers, each with this researcher as last author, pair carefully engineered evaluation protocols — typed decision readouts in one, rule-based verification in the other — with candid limitation notes about synthetic-only data and server-side heterogeneity.

Timeline

2026Publishes “Agents That Edit Documents” as last author, an agentic PDF-forgery benchmark reporting an 81.1% raw success rate (46.2% under strict filtering) against a non-agentic control.
2026Publishes “Open-Jev Judgments on CallScreenBench” as last author, showing JevLite typed-decision readouts on Qwen3-4B non-inferior to an LLM judge for scam-call screening at ~30x lower latency.
2026Both catalogued papers land on arXiv in the same year, splitting evenly between NLP and AI agent safety / document forensics — a recent senior-authorship output burst across two fields.

Papers in the Evolve catalog

Frequent co-authors

Simiao Ren×2 Ankit Raj×2 Tommy Duong×2 Yuxin Zhang×2 Dennis Ng×2 Xingyu Shen×2 Kidus Zewde×2 Yuchen Zhou×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 not available; the profile is name-resolved from public bibliographic data.
All bibliometric scores derive from Evolve's catalog of only two 2026 arXiv preprints; no venue or peer-review history is included.
Output, impact, and activity dimensions lack sufficient data and are intentionally left unscored; no citation counts or h-index were available.
Both papers self-report limitations — fully synthetic evaluation data and server-side heterogeneity — which bound what the reported benchmark numbers can support.