Audio-language models for telecom fraud detection — FRAUDSkill skill optimization and a refreshable audio benchmark
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-19) 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—
No publication-volume metric is available in the current snapshot, so the output axis is not scored. It would measure total papers, yearly production, and author-position distribution, which require a fuller bibliographic record than the two reviewed papers.
Impact—
Impact is not yet scored because no citation counts, h-index, or venue-level influence data appear in the snapshot. The axis would measure how widely the two fraud-detection papers are cited and adopted by other benchmark and method work.
Focus0/100
The radar scores focus at 0/100, computed from a field distribution split evenly at 50/50 between two nearly identical speech-and-audio labels — audio-language model benchmarking for fraud detection and audio-language models for fraud detection. Within the two reviewed papers, topical coverage is therefore highly concentrated on one applied problem, telecom fraud detection with audio-language models, with no recorded diversification into other fields.
Collaboration58/100
Collaboration scores 58/100, based on 17 distinct co-authors in the record. Twelve recur as frequent collaborators, including Huiyuan Liu, Shun Zhang, Qifan Wang, and Yifan Wang. Author positions differ by project: first author on the FRAUDSkill method paper and middle author on the TeleAntiFraud 2.0 benchmark, suggesting contribution within a larger, stable team rather than solo work.
Activity—
Research activity is not scored: the snapshot lacks a per-year publication series beyond the two 2026 arXiv papers. This axis would measure publication cadence and sustained research tempo over multiple years, data a longer publication history would provide.
Reproducibility78/100
Reproducibility scores 78/100, a mean taken across two reviewed papers. The TeleAntiFraud 2.0 benchmark is rated high readiness, supported by its refreshable design and collapse-aware analysis, while FRAUDSkill is rated medium. Evolve's review of FRAUDSkill notes that ablations attribute most of the +31.96 Macro-F1 gain to structured post-processing rather than the proposed skill search, which affects how cleanly the method itself reproduces.
Research summary
Chengxian Hu is an early-stage researcher whose catalogued work centers on audio-language models for telecom fraud detection. The Evolve record contains two 2026 arXiv papers that form a matched pair: TeleAntiFraud 2.0, a refreshable, profile-grounded, audio-based benchmark introducing near-domain negatives and collapse-aware analysis, and FRAUDSkill, a first-author method paper applying structured frozen-weight skill optimization to audio anti-fraud detection on the TeleAntiFraud benchmark. FRAUDSkill reports a +31.96 Macro-F1 gain, though Evolve's review cautions that ablations attribute most of that gain to structured post-processing rather than the skill search itself. The field distribution is an even 50/50 split between two closely overlapping speech-and-audio labels, indicating a tightly focused applied portfolio. Collaboration is moderate (58/100), with 17 distinct co-authors recorded and twelve frequent collaborators including Huiyuan Liu, Shun Zhang, and Qifan Wang; authorship spans first and middle positions, consistent with work inside a larger team. Reproducibility is comparatively strong (78/100 across the two papers): the benchmark is rated high readiness while the method paper is rated medium. Output, impact, and activity axes lack sufficient data, and no institution, country, or ORCID is attached, so this profile is name-resolved from public bibliographic data.
Collaboration pattern
The record lists 17 distinct co-authors and a collaboration score of 58/100. Twelve collaborators recur, including Huiyuan Liu, Shun Zhang, Qifan Wang, Yifan Wang, Yijin Zhou, and Yuxi Zhao. Authorship roles shift across the two 2026 papers — first author on FRAUDSkill, middle author on TeleAntiFraud 2.0 — indicating a core team producing both a shared benchmark and a method built on it.
Research keywords
Highlights
FRAUDSkill: first-author method for audio anti-fraud detection
TeleAntiFraud 2.0: a refreshable audio benchmark for telecom fraud
A matched benchmark-plus-method pairing
The two 2026 papers form one coherent line: a refreshable fraud-detection benchmark and a frozen-weight skill-optimization method evaluated on it, giving this early-stage profile an unusually integrated footprint with a mean reproducibility score of 78/100 across the pair.
Timeline
- 2026TeleAntiFraud 2.0 released on arXiv: a refreshable, profile-grounded, audio-based benchmark for telecom fraud detection with near-domain negatives.
- 2026FRAUDSkill posted as a first-author paper: structured frozen-weight skill optimization reporting +31.96 Macro-F1 on the TeleAntiFraud benchmark.
- 2026The benchmark and method papers appear together, anchoring this profile in audio-language models for telecom fraud detection with 17 recorded co-authors.
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.
- This profile is name-resolved from public bibliographic data; no institution, country, or ORCID is attached to the record.
- Radar scores and co-author counts come from public bibliometric data; only the two reviewed arXiv papers underlie the qualitative assessments.
- Output, impact, and activity dimensions are unscored because citation counts and a full publication history are absent from the snapshot.
- Evolve verdicts are reviewer assessments: the +31.96 Macro-F1 attribution for FRAUDSkill and the synthetic-only data limitation of TeleAntiFraud 2.0 come from those reviews, not independent replication.