Mohan Shi
University of California, Los Angeles · United States
UCLA researcher advancing parameter-efficient adaptation and hallucination mitigation in speech foundation models
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
6
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 — this axis would measure publication volume and productivity trajectory over time. The snapshot provides yearly work counts and an h-index of 6, but no composite output score has been computed, so it is left unscored here rather than estimated.
Impact—
Not yet scored — this axis would capture citation influence and scholarly reach accumulated over a career. Beyond a raw h-index of 6 and i10-index of 3, no citation totals were provided, so the data is insufficient and no impact estimate is offered.
Focus0/100
The radar records a focus score of 0, yet the raw topic counts point the other way: 31 counts in Speech Recognition and Synthesis, 29 in Speech and Audio Processing, and 17 in Speech and Dialogue Systems dominate, and fieldDist splits 0.5/0.5 entirely within speech processing. Taken together the work is narrowly specialized in speech and ASR; the zero score appears to be a scoring artifact rather than genuine scatter.
Collaboration51/100
Scored 51 from real bibliometric data: 12 distinct coauthors over the trailing three years, with recurring partners including Natarajan Balaji Shankar, Zilai Wang, Zihan Wang, Kaiyuan Zhang, Abeer Alwan, and Eray Eren. Both reviewed 2026 papers are co-authored — first author on one, middle author on the other, neither corresponding — fitting a steady contributor in a compact, stable team.
Activity89/100
Scored 89 on real data: recent3yWorks stands at 20, and countsByYear climbs from 1 work per year in 2018–2020 to 3 in 2022, 9 in 2023, 4 in 2024, 6 in 2025, and 10 in 2026 — the highest on record — with 2021 the only gap. The trajectory shows strongly accelerating, sustained output concentrated in the most recent three years.
Reproducibility78/100
Scored 78, matching a reproMean of 78 across 2 repro-scored papers. AURA is rated high reproducibility readiness, while Encoder Awakening via Adapters is rated medium — Evolve judges it convincing and well-ablated but flags thin statistical reporting and modest MyST gains. Both works expose concrete, checkable recipes, supporting an above-average reproducibility reading.
Research summary
Mohan Shi is a speech-processing researcher at the University of California, Los Angeles; this profile is name-resolved from OpenAlex bibliographic data. The indexed record runs from 2018 onward and accelerates sharply: 20 works in the trailing three years, with 10 works recorded in 2026 alone. Topic assignments concentrate heavily on speech — 31 counts in Speech Recognition and Synthesis, 29 in Speech and Audio Processing, and 17 in Speech and Dialogue Systems — with smaller footprints in Music and Audio Processing (5) and Domain Adaptation and Few-Shot Learning (4). Two 2026 arXiv papers anchor the Evolve catalog. "Encoder Awakening via Adapters", where he is first author, proposes two-stage encoder-adapter awakening followed by joint fine-tuning for domain-shifted ASR with Speech-LLMs, judged convincing and well-ablated. "AURA", where he is a middle author, introduces uncertainty-routed activation editing that reportedly cuts Whisper non-speech hallucination from 89.2% to 1.9% with roughly 500x fewer parameters than LoRA. Both works center parameter-efficient fine-tuning, adapters, LoRA, and cross-attention. Radar scores show high activity (89), solid reproducibility (78, averaged over two repro-scored papers), and moderate collaboration (51, with 12 distinct recent coauthors); the focus axis reads 0 despite tightly clustered topics. Output and impact are unscored, with a raw h-index of 6 and i10-index of 3.
Collaboration pattern
Collaboration is real but compact: the radar scores it 51, with 12 distinct coauthors over the trailing three years. Recurring partners in the record include Natarajan Balaji Shankar, Zilai Wang, Zihan Wang, Kaiyuan Zhang, Abeer Alwan, and Eray Eren. On the two reviewed 2026 papers he is first author once and middle author once, neither corresponding — a steady contributor inside a small, stable team rather than a lone author or a sprawling consortium.
Research keywords
Highlights
AURA: uncertainty-routed activation editing curbs Whisper hallucinations
First-author Encoder Awakening adapts Speech-LLMs to domain shift
Above-average reproducibility signal
The reproMean is 78 across 2 repro-scored papers: AURA is rated high readiness and Encoder Awakening medium. Both publish concrete, checkable recipes — adapters, LoRA, and activation-head edits — on widely used speech models such as Whisper.
Timeline
- 2018First indexed works appear in the OpenAlex record, with 1 work recorded in each of 2018, 2019, and 2020.
- 2023Yearly output jumps to 9 works — the highest count before 2026 — after the single gap year of 2021.
- 2024A three-year run begins that totals 20 works: 4 in 2024, 6 in 2025, and 10 in 2026.
- 2026Two Evolve-reviewed arXiv papers appear — first-author "Encoder Awakening via Adapters" and middle-author "AURA" — alongside a record 10 works in the year.
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.
- Bibliometrics are name-resolved from public OpenAlex data only; no ORCID was provided, so works by others with the same name may be mixed in.
- Output and impact radar axes are unscored; no citation totals beyond h-index 6 and i10-index 3 were provided.
- The recorded focus score of 0 conflicts with the tightly clustered speech topics and may be a scoring artifact rather than genuine breadth.
- Qualitative verdicts rest on only two Evolve-reviewed papers, both 2026 arXiv preprints; they are Evolve assessments, not peer review.