Jian Liang

Chinese Academy of Sciences · Australia

Unclaimed · AI-structured profileORCID OpenAlex

Machine learning researcher focused on confidence calibration for vision-language models and LLM-based agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Output (publication volume and productivity trajectory over time) is not scored: the underlying per-year, venue-normalized output data is insufficient in this snapshot, so no estimate is made despite the 16 works logged for 2026.

Impact—

Impact (citation uptake, h-index-based influence, and reach beyond the immediate team) cannot be scored: citation data in this snapshot is too thin — an h-index of 1 and i10-index of 1 — so no estimate is offered.

Focus100/100

Focus scores 100 because the field distribution in this snapshot collapses to a single field — machine learning / computer vision with a vision-language-model calibration emphasis (1 paper) — and the two Evolve-reviewed 2026 papers both target confidence calibration, one for CLIP-style vision-language models and one for LLM-based deep research agents. Topic counts (domain adaptation and few-shot learning 10; topic modeling 9) suggest broader underlying interests.

Collaboration46/100

Collaboration scores 46, a moderate level. The snapshot records 9 distinct coauthors, with recurring partners including Yuwei Liang, Yinuo Xu, Jianjie Cheng, Meng Wang, Yongcan Yu, Shuo Lu, Dapeng Hu, and Ran He. Both reviewed 2026 papers are multi-author arXiv works in which Jian Liang holds middle and last author positions respectively, consistent with contributing within a standing team rather than leading every effort.

Activity86/100

Activity scores 86, reflecting a steep recent ramp: the snapshot logs 18 works in the last three years, and counts by year show 1 in 2022, 1 in 2024, 1 in 2025, then 16 in 2026 — a concentration that drives the high score. Both Evolve-reviewed papers appeared in 2026 on arXiv, marking 2026 as the visible breakout year in this record.

Reproducibility100/100

Reproducibility scores 100, computed as a mean over only 1 scored paper. The temperature-scaling paper for test-time prompt tuning carries high reproducibility readiness, consistent with lightweight post-hoc calibration; the DualStake paper is rated medium, with single-seed reinforcement learning results limiting certainty. With one paper in the sample, the perfect score is promising but not yet robust evidence.

Research summary

Jian Liang (Chinese Academy of Sciences; ORCID 0000-0002-4243-0784) is an early-stage machine learning researcher whose visible work centers on confidence calibration. The bibliometric snapshot, name-resolved from OpenAlex, credits 18 works in the last three years, with a pronounced burst of 16 publications logged for 2026 — an activity score of 86 — against a still-thin citation footprint (h-index 1, i10-index 1). Topic counts show breadth across domain adaptation and few-shot learning (10), topic modeling (9), multimodal machine learning applications (5), advanced neural network applications (4), and natural language processing techniques (4), while the reviewed catalog concentrates in a single field, yielding a perfect focus score of 100. Two Evolve-reviewed 2026 arXiv papers define the current agenda: temperature scaling for calibrating test-time prompt tuning in CLIP-style vision-language models (high reproducibility readiness), and DualStake's dual-path confidence calibration for deep research agents via margin-clipped stake rewards (medium readiness; single-seed RL). The reproducibility mean is 100 over one scored paper. Collaboration is moderate (score 46, 9 distinct coauthors), anchored by partners such as Yuwei Liang, Yinuo Xu, and Ran He.

Collaboration pattern

Collaboration is moderate (score 46): the snapshot records 9 distinct coauthors, and the frequent-coauthor list is dominated by a stable cluster — Yuwei Liang, Yinuo Xu, Jianjie Cheng, Meng Wang, Yongcan Yu, Shuo Lu, Dapeng Hu, and Ran He — suggesting work within one or two standing groups rather than a wide transient network. On the two reviewed 2026 papers Jian Liang holds middle and last author positions, neither as corresponding author.

Research keywords

confidence calibrationtest-time prompt tuningtemperature scalingvision-language modelsCLIPexpected calibration errordeep research agentsverbalized confidenceGRPO reinforcement learningretrieval-augmented generationdomain adaptation and few-shot learningtopic modeling

Highlights

Zero-shot-anchored temperature scaling calibrates test-time prompt tuning
In this 2026 arXiv paper (middle author), CoTS/E-CoTS cut TPT's calibration error on CLIP-style vision-language models by anchoring temperature scaling to zero-shot logits while preserving accuracy; Evolve notes the evidence is broad, though ECE largely mirrors the zero-shot anchor. Reproducibility readiness: high.
DualStake aligns evidence and answer confidence in deep research agents
This 2026 arXiv paper (last author) proposes margin-clipped stake rewards that align verbalized evidence and answer confidence with correctness in deep research agents, drawing on GRPO reinforcement learning and retrieval-augmented generation; gains are broad, but single-seed RL results limit certainty. Reproducibility readiness: medium.
A 2026 output burst defines the visible record
Counts by year show 1, 1, 1 works in 2022, 2024, and 2025, then 16 in 2026, with 18 works across the last three years — the OpenAlex-visible record is dominated by a single breakout year, while the citation footprint (h-index 1, i10-index 1) has yet to catch up.

Timeline

2022First visible work in the snapshot: the OpenAlex counts-by-year series logs 1 publication for 2022.
2024Output remains sparse: 1 publication logged for 2024.
2025Another quiet year with 1 logged publication, immediately before the 2026 ramp-up.
2026Breakout year: 16 works logged, including two Evolve-reviewed arXiv papers on confidence calibration — temperature scaling for test-time prompt tuning and DualStake for deep research agents.

Papers in the Evolve catalog

Frequent co-authors

Yuwei Liang×2 Yinuo Xu×1 Jianjie Cheng×1 Meng Wang×1 Yongcan Yu×1 Shuo Lu×1 Dapeng Hu×1 Yinuo Xu×1 Ran He×1

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 snapshot is name-resolved from OpenAlex; 'Jian Liang' is a common name, and record coverage may be incomplete or mixed.
Output and impact dimensions are unscored because underlying data is missing; h-index (1) and i10-index (1) come from this snapshot only.
Qualitative assessment rests on only 2 Evolve-reviewed papers (both 2026 arXiv preprints); the reproducibility mean is based on a single scored paper.
Affiliation metadata lists 'Chinese Academy of Sciences' with country code 'AU'; the affiliation-country pairing may be inconsistent in the source record.