University of Macau researcher on efficient LLM alignment: neuron-guided fine-tuning and multilingual reward modeling
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
The publication-output axis is unscored: this snapshot lacks the volume metrics, such as cumulative works and output trajectory, needed to compute it. It would measure how productive the researcher is over time relative to field norms.
Impact—
The impact axis is unscored: h-index and i10-index are absent (recorded as 0) and no citation counts are available in this OpenAlex snapshot. It would measure citation-based influence and diffusion of the researcher's publications.
Focus0/100
The measured focus score is 0, and the raw distribution explains it: work splits exactly evenly (0.5/0.5) between LLM fine-tuning and reward modeling, while OpenAlex topics scatter across topic modeling (3 papers), NLP techniques (2), and three single-paper topics. This reads as breadth across adjacent alignment subfields rather than scattered output, but the even split keeps the axis at the bottom of the scale.
Collaboration61/100
The collaboration score is 61, grounded in 20 distinct co-authors over 7 works in the recent three-year window. Twelve recurring names appear, including Junchao Wu, Yichao Du, Longyue Wang, Weihua Luo, Zeyu Wu, Shudong Liu, Runzhe Zhan, Xin Chen, Shu Yang, Jinsong Su, Peng Lai, and Weibo Gao. On reviewed papers Wong sits in middle and last author positions, with no corresponding-author role recorded.
Activity61/100
The activity score of 61 reflects a sustained recent rhythm: OpenAlex counts 7 works in the recent three-year window, with 4 works in 2025 and 3 in 2026. All three catalog papers are 2026 arXiv preprints, indicating a preprint-first publishing pattern concentrated in the current period rather than a long cumulative record.
Reproducibility55/100
The reproducibility score is a mean of 55 across 2 assessed papers, both rated medium for repro-readiness. Evolve's verdicts identify concrete replication gaps: NGFT reports broad gains without significance testing, and UniRRM lacks significance tests while carrying an LLM-verified-data circularity risk in its training pipeline.
Research summary
Derek F. Wong is a researcher at the University of Macau whose recent work concentrates on large language model alignment, splitting evenly between LLM fine-tuning and reward modeling in the recorded field distribution (0.5/0.5). The Evolve catalog reviews three 2026 arXiv preprints, two with full editorial verdicts. Neuron-Guided Fine-Tuning (NGFT), with Wong as last author, uses neuron activations to unify instruction-data selection, sparse updates, and anti-forgetting in supervised fine-tuning; Evolve notes broad reported gains but no significance testing. UniRRM, with Wong as a middle author, presents a unified reasoning reward model plus MixReward, a 64.5K-preference dataset spanning 103 languages that unifies pairwise, listwise, and point-wise judging. A third preprint, IndicDetect, evaluates cross-lingual detection of LLM-generated text for Hindi, Telugu, and Tamil. OpenAlex bibliometrics show 7 works in the recent three-year window (4 counted in 2025, 3 in 2026), 20 distinct co-authors, and a reproducibility mean of 55 across 2 assessed papers. Citation indicators are absent from this snapshot, so output and impact axes remain unscored; the profile is name-resolved from public bibliographic data.
Collaboration pattern
Wong's collaboration footprint is broad rather than deep: 20 distinct co-authors across 7 works in the recent three-year window, yielding a collaboration score of 61. Twelve names recur, including Junchao Wu, Yichao Du, Longyue Wang, Weihua Luo, Zeyu Wu, Shudong Liu, Runzhe Zhan, Xin Chen, Shu Yang, Jinsong Su, Peng Lai, and Weibo Gao. On the reviewed 2026 papers he appears as middle or last author, with no corresponding-author role recorded.
Research keywords
Highlights
Neuron-guided fine-tuning unifies efficient LLM alignment
UniRRM and MixReward unify multilingual reward models
Benchmarking cross-lingual detection of LLM-generated text
IndicDetect (2026, arXiv, last author position) evaluates cross-lingual detection of LLM-generated text for Hindi, Telugu, and Tamil. This paper has not yet received an Evolve editorial verdict, keywords, or a repro-readiness rating.
Timeline
- 2025OpenAlex records 4 works in this year within the available snapshot window, the higher of the two counted years.
- 2026NGFT preprint appears on arXiv, proposing neuron-guided fine-tuning for efficient alignment; Wong is last author.
- 2026UniRRM preprint appears on arXiv, introducing a unified multilingual reasoning reward model with the 103-language MixReward dataset; Wong is a middle author.
- 2026IndicDetect preprint appears on arXiv, evaluating cross-lingual LLM-generated text detection for Hindi, Telugu, and Tamil; Wong is last author.
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.
- Bibliometric indicators are computed from OpenAlex only; no ORCID was provided, so this profile is name-resolved from public bibliographic data.
- Citation-impact indicators (h-index, i10-index) and the output and impact dimensions are unavailable in this snapshot and are left unscored rather than estimated.
- All catalog papers are 2026 arXiv preprints; Evolve editorial verdicts exist for two of the three, and IndicDetect carries no verdict, keywords, or repro-readiness rating.
- Reproducibility and activity scores rest on a small base of 2 assessed papers and 7 recent works, so scores may shift as data accrues.