University of Queensland researcher bridging speech processing and LLM interpretability via activation editing
2
papers here
2
reviewed by Evolve
0
first-author
0
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—
Insufficient data: this axis measures publication volume — total papers, annual output, and authorship roles across venues. Only two 2026 arXiv preprints are cataloged for this researcher, so no output score can be computed yet.
Impact—
Insufficient data: this axis measures citation-based influence — citation counts, aggregate indices, and uptake of the two preprints. No citation statistics exist in the provided snapshot, so no impact score is estimated.
Focus0/100
Focus scores 0, reflecting a perfectly even split in the field distribution: 50% speech processing / automatic speech recognition and 50% machine learning interpretability for NLP. Each of the two cataloged papers sits in a different field, so no single area dominates the record. This breadth connects activation-editing methods in speech foundation models with probing and steering analysis of language models, but research concentration is minimal so far.
Collaboration55/100
Collaboration scores 55, based on 15 distinct co-authors recorded across the reviewed record. The frequent-coauthor list shows repeats — Natarajan Balaji Shankar and Mohan Shi each appear twice — suggesting at least two recurring working ties. Other named collaborators include Zilai Wang, Kaiyuan Zhang, Abeer Alwan, Manjiang Yu, Hongji Li, Junwei Chen, and Xue Li, spanning both fields. The moderate score fits a mid-sized, partly recurring network.
Activity—
Insufficient data: this axis measures publishing cadence — papers per year, recency of the latest output, and sustained posting over time. The snapshot holds two preprints both dated 2026 with no multi-year history, so no activity score is computed.
Reproducibility100/100
Reproducibility is the strongest quantified axis: the dimension scores 100, with a reproMean of 100 across 2 repro-scored papers, and both cataloged preprints carry high reproducibility readiness. AURA backs its central claim with a concrete figure — Whisper non-speech hallucination falling from 89.2% to 1.9% with about 500x fewer parameters than LoRA — while the answer-basin paper's probing and steering experiments are rated suggestive but correlational, yet still high on reproducibility.
Research summary
Zihan Wang is a researcher at The University of Queensland whose cataloged work sits at the intersection of speech technology and language-model interpretability. The bibliometric field distribution splits evenly — 50% speech processing / automatic speech recognition and 50% machine learning interpretability for NLP — and the Evolve catalog holds two 2026 arXiv preprints on which Wang appears as a middle-position, non-corresponding co-author. The first, AURA, applies uncertainty-routed activation editing to acoustically ground speech foundation models: Evolve's review reports a reduction of Whisper's non-speech hallucination rate from 89.2% to 1.9% with roughly 500x fewer trainable parameters than LoRA, while noting results are Whisper-only. The second, the Answer-Basin Representation Hypothesis, argues that linear directions in language models are organized by answer-mass basins rather than discrete concept labels; the verdict calls the probing and steering evidence suggestive but correlational and notes the title overclaims. Both papers rate high on reproducibility readiness, and the reproducibility dimension scores 100 (reproMean across 2 papers). Collaboration scores 55 over 15 distinct co-authors, with repeats including Natarajan Balaji Shankar and Mohan Shi. Output, impact, and activity lack sufficient data and are left unscored.
Collaboration pattern
The collaboration dimension scores 55, with 15 distinct co-authors on record. Two names repeat in the frequent-coauthor list — Natarajan Balaji Shankar and Mohan Shi — pointing to recurring working relationships, while the remaining network (Zilai Wang, Kaiyuan Zhang, Abeer Alwan, Manjiang Yu, Hongji Li, Junwei Chen, Xue Li) spans both fields. Both cataloged papers are middle-position, non-corresponding co-authorships.
Research keywords
Highlights
AURA cuts Whisper non-speech hallucination from 89.2% to 1.9%
Argues LM linear directions encode answer-mass basins, not concept labels
Timeline
- 2026Co-authored (middle position) AURA on arXiv: uncertainty-routed activation editing for acoustic grounding in speech foundation models.
- 2026Co-authored (middle position) the Answer-Basin Representation Hypothesis preprint on arXiv, questioning concept-based probing and steering claims in language models.
- 2026Both 2026 preprints entered the Evolve review catalog with high reproducibility readiness; the reproducibility dimension scored 100 across the pair (reproMean 100).
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.
- Dimension scores and the co-author count come from public bibliometric snapshots only; no citation, h-index, or venue-level metrics are included in this input.
- Zihan Wang is a common name; this profile is name-resolved from public bibliographic data linked to ORCID 0009-0003-1583-434X, and 'Zihan Wang' itself appears in the frequent-coauthor list, suggesting possible disambiguation noise.
- The reviewed catalog contains only two papers — both 2026 arXiv preprints with middle-position, non-corresponding authorship — so output, impact, and activity remain unscored.
- Both entries carry Evolve review verdicts on preprints; no peer-reviewed venue, acceptance, or publication history is present in the data.