Early-career CAS researcher building confidence calibration for vision-language models and LLM agents
3
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
3
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 output axis measures publication volume and productivity trends. A score is not yet available: the OpenAlex record is sparse (nine indexed works, several likely name-resolved), so reliable output volume cannot be assessed from current data.
Impact—
The impact axis measures citation influence — total citations, h-index trajectory, and downstream uptake. Insufficient data exist to score it: the snapshot lists only h-index 3 and i10-index 1, too thin and recent to support a meaningful estimate.
Focus0/100
Grounded in raw metrics, focus is dispersed: the field distribution splits evenly at 0.5/0.5 between computer vision/machine learning and vision-language-model calibration, and OpenAlex topics scatter across five clusters — advanced neural network applications (3), robot manipulation and learning (3), face and expression recognition (2), reinforcement learning in robotics (2), and domain adaptation and few-shot learning (2). The 0 focus score reflects this breadth.
Collaboration51/100
The collaboration score of 51 is mid-range. The OpenAlex record lists 12 distinct coauthors; recurring partners include Ran He, Jian Liang, Tieniu Tan, Siru Jiang, Meng Wang, and Yinuo Xu. All three reviewed 2026 papers are collaborative arXiv works — first author on the temperature-scaling study, middle author on the other two.
Activity47/100
The activity score of 47 rests on 4 works in the last three years and a 2026 burst of 3 works — the busiest year in countsByYear — after sparse earlier output (single works in 2017, 2019, 2020, and 2024; two in 2021). Recent momentum contrasts with a historically thin publication cadence.
Reproducibility100/100
The reproducibility score of 100 (reproMean across 2 assessed papers) is the profile's strongest signal: both vision-language calibration papers carry high reproducibility readiness, while DualStake is rated medium, with the review noting single-seed reinforcement-learning results as a limit on certainty.
Research summary
Yuwei Liang is a researcher at the Chinese Academy of Sciences whose recent work centers on confidence calibration for foundation models, spanning vision-language models, test-time adaptation, and LLM-based agents. The Evolve-reviewed catalog holds three 2026 arXiv preprints: a first-author study applying zero-shot-anchored temperature scaling to calibrate test-time prompt tuning for CLIP-style models; a co-authored selective test-time adaptation method that skips unreliable per-sample updates using cross-augmentation agreement; and DualStake, which aligns evidence- and answer-level confidence in deep research agents via margin-clipped stake rewards under reinforcement learning. Bibliometric data under an OpenAlex identity record nine indexed works between 2017 and 2026, four in the last three years, twelve distinct coauthors, an h-index of 3, and an i10-index of 1 — an early-stage record that is assembling quickly around calibration. Reproducibility is a clear strength: both assessed vision-language papers carry high reproducibility readiness, yielding a reproMean of 100, while the agent paper is rated medium. The broader OpenAlex topic profile also lists robotics, face and expression recognition, and few-shot learning, which may partly reflect name resolution on a relatively common name.
Collaboration pattern
Bibliometric records list 12 distinct coauthors, with the most frequent partners including Ran He, Jian Liang, Tieniu Tan, Siru Jiang, Meng Wang, and Yinuo Xu — a network anchored at the Chinese Academy of Sciences. In the reviewed 2026 work, Yuwei Liang leads as first author on the temperature-scaling calibration study and contributes as a middle author on selective adaptation and DualStake, suggesting a mix of leading and supporting roles within small teams.
Research keywords
Highlights
First-author temperature scaling for calibrated test-time prompt tuning
Selective adaptation: knowing when not to adapt vision-language models
DualStake: dual-path confidence calibration for deep research agents
Timeline
- 2017Earliest year in the OpenAlex counts-by-year record, with one indexed work.
- 2021Two indexed works — the peak of the early period in the bibliometric record.
- 2024One indexed work in the most recent pre-2026 year on record.
- 2026Most productive year with three works, covering all three Evolve-reviewed papers: first-author temperature scaling for test-time prompt tuning, selective adaptation for vision-language models, and DualStake confidence calibration for deep research agents.
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 OpenAlex under a name-resolved identity; Yuwei Liang is a relatively common name, so earlier records may include name-resolution noise.
- Output and impact axes are unscored due to insufficient data; only h-index 3 and i10-index 1 appear in the raw snapshot.
- The topic profile (robot manipulation, face and expression recognition, few-shot learning) is broader than the reviewed 2026 calibration papers and may reflect name-resolution ambiguity.
- The reproducibility score of 100 is computed from only 2 assessed papers, and all reviewed items are 2026 arXiv preprints with no peer-review venue information in scope.