UCLA researcher spanning speech AI and LLM-agent security: Whisper hallucination control to Z3-verified agent policies
4
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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-volume metrics such as total output, venue breakdown, and per-year counts are not yet available in this snapshot, so productivity cannot be scored; this axis would measure aggregate research output over time.
Impact—
No citation counts, h-index, or other influence indicators are available yet; this axis would measure citations to the reviewed papers and their uptake in follow-up work.
Focus0/100
The radar focus score is 0, and the raw field distribution explains why: exactly 50% speech processing / automatic speech recognition and 50% AI security / systems security (LLM agent defense). This is a deliberately two-front portfolio — hallucination control and parameter-efficient speech adaptation on one side, agent-action governance, membership inference, and PII detection on the other — rather than a single narrowly specialized line.
Collaboration63/100
The collaboration score is 63, computed from 23 distinct co-authors recorded in the bibliometric data. Frequent collaborators include Abeer Alwan, Chuan Wang, Joey Zhong, Paul Fryzel, Kyle Polley, Jerry Ma, Ninghui Li, Kunlin Cai, Zihang Xiang, Jinghuai Zhang, Fnu Suya, and Yuan Tian — a roster consistent with the recorded 50/50 split between the speech and security fields.
Activity—
Temporal activity data is insufficient — all four reviewed papers are 2026 arXiv preprints, leaving no multi-year record; this axis would measure publishing cadence and continuity over a longer window.
Reproducibility78/100
The reproducibility score is 78, the mean across 2 scored papers. Within the reviewed catalog, AURA is rated high reproducibility readiness, while ActGov, Hearing the Whispers, and PII-TRACE are rated medium — a profile that leans toward reproducible, artifact-friendly work.
Research summary
Kaiyuan Zhang is a researcher at the University of California, Los Angeles whose reviewed work spans two roughly equal fronts: speech processing / automatic speech recognition, and AI / systems security centered on LLM agent defense. The Evolve catalog holds four 2026 arXiv papers, two of them first-authored. On the speech side, AURA applies uncertainty-routed activation editing to Whisper, cutting non-speech hallucination from 89.2% to 1.9% with roughly 500x fewer parameters than LoRA, while Hearing the Whispers reports the first black-box membership inference attacks on fine-tuned TTS models (AUC above 0.8). On the security side, the first-authored ActGov governs LLM agent actions with LLM-drafted, Z3-verified policies, driving indirect-prompt-injection attack success to roughly zero on AgentDojo/AgentDyn, and PII-TRACE contributes a 13,000-dialogue multilingual benchmark for cross-turn PII detection plus a 0.6B-parameter detector. Bibliometrically, the profile records 23 distinct co-authors, an even 50/50 split across the two fields, and a mean reproducibility score of 78 across two scored papers. Caveats: the evidence base is limited to these reviewed papers, and output, impact, and activity metrics are not yet available.
Collaboration pattern
The recorded network spans 23 distinct co-authors, with frequent collaborators including Abeer Alwan, Chuan Wang, Joey Zhong, Paul Fryzel, Kyle Polley, Jerry Ma, Ninghui Li, Kunlin Cai, Zihang Xiang, Jinghuai Zhang, Fnu Suya, and Yuan Tian. The even 50/50 split between speech processing and AI security suggests collaboration across two communities, though the data does not map individual co-authors to specific papers.
Research keywords
Highlights
ActGov drives indirect prompt injection attack success to ~0 on LLM agents
AURA cuts Whisper non-speech hallucination from 89.2% to 1.9%
First black-box membership inference attacks on fine-tuned TTS models
PII-TRACE: first cross-turn PII benchmark for multi-turn LLM conversations
Timeline
- 2026Publishes ActGov (first author), governing LLM agent actions with policy-constrained validation and driving IPI attack success to ~0 on AgentDojo/AgentDyn.
- 2026Publishes PII-TRACE (first author), the first cross-turn PII detection benchmark with 13,000 multilingual dialogues and a 0.6B detector.
- 2026Publishes AURA on uncertainty-routed activation editing, cutting Whisper non-speech hallucination from 89.2% to 1.9% with ~500x fewer parameters than LoRA.
- 2026Publishes Hearing the Whispers, the first black-box membership inference attacks on fine-tuned TTS models, with AUC above 0.8.
Papers in the Evolve catalog
- first author
ActGov: Governing LLM Agent Actions via Policy-Constrained Validation
AURA: Uncertainty-Routed Activation Editing for Acoustic Grounding in Speech Foundation Models
Hearing the Whispers: Black-Box Membership Inference Attacks on Finetuned TTS Models
- first author
PII-TRACE: A Benchmark for Context-Aware PII Detection in Multi-Turn LLM Conversations
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.
- Bibliometrics (co-author counts, field distribution, radar scores) are derived from OpenAlex-based data only; no citation counts, h-index, or venue-tier metrics were provided.
- Evidence is limited to four 2026 arXiv preprints reviewed by Evolve; earlier or other-venue publications may be missing from this record.
- Output, impact, and activity dimensions are unscored because the underlying data is insufficient.
- Country and ORCID are not recorded; identity is name-resolved from public bibliographic data.