NLP researcher at the University of Turku: LLM distillation and test-time RL for efficient adaptation
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
21
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 scored yet: this OpenAlex record has no verified complete works list or venue breakdown attached, so total publication volume cannot be computed. This axis would measure cumulative output and its split across years and publication venues.
Impact—
Not scored yet: citation linkage in this snapshot is insufficient for reliable aggregation, even though an h-index of 21 is listed. This axis would measure citation-based influence, citing-work growth and uptake of the researcher's methods.
Focus0/100
Focus scores 0, the lowest radar value, reflecting a deliberately broad portfolio rather than a single niche: the field distribution splits exactly evenly (0.5/0.5) between LLM training & distillation and LLM efficient adaptation / test-time RL, while topic counts span topic modeling (32), NLP techniques (20), sentiment analysis (16), mental-health writing (15) and domain adaptation / few-shot learning (12).
Collaboration39/100
Collaboration scores 39 on a compact network: the snapshot lists only 6 distinct co-authors, with recurring partners Yilei Tu, Zihao Li, Jörg Tiedemann, Fei Yuan, Naveen Vakada and Mingyuan Li. In both Evolve-reviewed papers this researcher holds last- or middle-author positions without corresponding status, indicating a contributor role inside larger LLM-training project teams.
Activity98/100
Activity is the strongest axis at 98: the works-per-year series climbs from 2 works in 2018 into a sustained 9–12 range from 2021 onward — 12 in both 2021 and 2024, 9 in 2026 — with 28 works in the last three years, a consistently high and accelerating publication tempo.
Reproducibility78/100
Reproducibility scores 78, averaged over the 2 Evolve-reviewed papers. The specialist-distillation study is rated high readiness — an 8B-parameter, well-controlled comparison — while label-free steering is medium: honestly reported, but with train-on-eval and MMAU confounds flagged. The pattern suggests setup-aware, controlled experimental practice whose statistical rigor (single-run results) still has room to strengthen.
Research summary
Shaoxiong Ji is a researcher at the University of Turku, Finland, whose profile is resolved from OpenAlex bibliographic data. The work is rooted in natural language processing — topic modeling (32 works), NLP techniques (20), sentiment analysis (16), mental health via writing (15), and domain adaptation / few-shot learning (12) — while the current field distribution splits evenly between LLM training & distillation and LLM efficient adaptation / test-time RL. Two 2026 arXiv papers in the Evolve catalog capture the present agenda: one compresses test-time reinforcement learning into roughly 100K bias-only steering parameters that approach full fine-tuning scores on MATH-500, and the other shows that QA-only specialist tuning implicitly determines which reasoning trajectories distilled students inherit. Activity is the strongest radar axis at 98: output rose from 2 works in 2018 to a sustained 9–12 per year since 2021, with 28 works in the last three years. The h-index is 21 (i10-index 27), though citation-linked impact is not scored here. Collaboration is compact (score 39, 6 distinct co-authors, recurring partners including Jörg Tiedemann and Fei Yuan), and reproducibility averages 78 across the two reviewed papers. In both reviewed papers this researcher appears as last or middle author rather than corresponding author, consistent with senior-contributor roles in team-driven projects.
Collaboration pattern
Collaboration centers on a small, stable core of 6 distinct co-authors, giving a radar score of 39. Recurring partners include Yilei Tu, Zihao Li, Jörg Tiedemann, Fei Yuan, Naveen Vakada and Mingyuan Li. In the two Evolve-reviewed papers this researcher is neither first nor corresponding author — last on the steering paper, middle on the distillation paper — consistent with senior-contributor roles inside larger LLM-training project teams.
Research keywords
Highlights
Label-free steering compresses test-time RL into ~100K bias-only parameters
Specialist models trained without reasoning trajectories still steer students
Timeline
- 2018Earliest year in the works-per-year series, with 2 works recorded.
- 2021Output jumps to 12 works (from 5 in 2020), starting a sustained 9–12 works-per-year phase.
- 2024Second peak of the series, again reaching 12 works in a single year.
- 2026Both Evolve-catalog arXiv papers appear this year — label-free test-time RL steering and trajectory-free specialist distillation — within a yearly count of 9 works.
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 come from an OpenAlex snapshot (identity: openalex; ORCID 0000-0003-3281-8002) and were not independently verified.
- Output and impact radar axes are unscored because citation-linked records in the provided data are insufficient.
- Reproducibility scores and verdicts cover only the 2 papers reviewed by Evolve and may not represent the full portfolio.
- The timeline uses only years present in the provided papers and the counts-by-year series.