First-author researcher on how LLMs revise beliefs — dialogue clarification and agent social reasoning
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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 yet scorable: Evolve holds no publication-volume, venue-count, or authorship-share data for this author beyond the 2 catalog papers, so total scholarly output cannot be measured. This axis would track papers per year, venues, and first/corresponding-author share.
Impact—
Not yet scorable: no citation counts, citation velocity, or downstream-usage signals are available for the 2026 arXiv papers, so influence cannot be estimated. This axis would measure how often the work is cited and taken up by later research.
Focus0/100
Scores 0 on the 0-100 scale because the field distribution splits exactly 0.5/0.5 between natural language processing (LLM dialogue evaluation) and artificial intelligence (LLM agents). Across only two first-authored 2026 papers, this even split signals breadth rather than drift: each paper occupies one field, yet both target the same failure mode of LLMs not revising a committed belief or state.
Collaboration46/100
Scores 46, built on 9 distinct coauthors across the 2 catalog papers. The frequent-coauthor list shows a compact recurring core — Xiaolin Li and Fei Wang each recur most often, with Jubin Chheda, Robert Douglas, Rajeshkumar Golani, and Yunda Liu also appearing — consistent with a small, stable team rather than a wide rotating network.
Activity—
Not yet scorable: both catalog papers are dated 2026, leaving no multi-year record to compute a publication cadence or recent-engagement trend. This axis would track output rhythm and continuity of research activity over time.
Reproducibility38/100
Scores 38, the mean repro-readiness across 2 scored papers. The dialogue-evaluation paper is rated low, with Evolve flagging judge circularity and the absence of statistical significance tests; the ReAdapt agent paper is rated medium, limited to a single model with no ablation or statistical analysis. Both verdicts converge on missing statistical rigor as the key gap.
Research summary
Jianzhe Lin is an early-stage researcher whose Evolve profile is built on two first-authored 2026 arXiv papers that attack one underlying weakness in large language models: the difficulty of revising a belief once it has been formed. The first paper, in LLM dialogue evaluation, argues that models commit to early answers and then treat later user clarification as additional context rather than as correction. The second, in LLM agents, reports that ReAdapt's explicit social-state updates lift relational accuracy from 37% to 51% (and from 69% to 77%), which Evolve reads as a credible proof-of-concept limited by a single model and no ablations. Bibliometrically, the field distribution splits exactly 0.5/0.5 between natural language processing (LLM dialogue evaluation) and artificial intelligence (LLM agents), yielding a focus score of 0, though both papers share the same revision theme. Collaboration scores 46, with 9 distinct coauthors and a recurring core around Xiaolin Li and Fei Wang. Reproducibility averages 38 across the 2 scored papers, with both Evolve verdicts citing missing statistical tests. Output, impact, and activity axes remain unscored because volume, citation, and longitudinal data are not in scope. Identity is resolved from public bibliographic data via ORCID 0000-0002-1559-1048; institution and country are unrecorded.
Collaboration pattern
Collaboration scores 46. Across the 2 catalog papers Jianzhe Lin worked with 9 distinct coauthors, and the frequent-coauthor list points to a stable core: Xiaolin Li and Fei Wang recur most often, joined by Jubin Chheda, Robert Douglas, Rajeshkumar Golani, and Yunda Liu. Lin takes first-author position on both papers, with no corresponding-author roles recorded, consistent with driving the research and writing within one small, recurring team.
Research keywords
Highlights
Committing early: clarifications read as context, not correction
ReAdapt lifts relational social reasoning from 37% to 51%
One failure mode across two fields: hard to let go
Both catalog papers converge on the same phenomenon: once an LLM forms a hypothesis or social state, it resists updating it — whether in multi-turn dialogue facing user clarifications or in agent benchmarks requiring relational state revision. This gives the early-stage profile a coherent identity despite the 0.5/0.5 split across two fields.
Timeline
- 2026Publishes the first-authored arXiv study 'Clarification Is Not Correction: LLMs Fail to Let Go' in LLM dialogue evaluation.
- 2026Publishes the first-authored arXiv paper 'When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning' in LLM agents.
- 2026Both papers enter the Evolve catalog in the same year, establishing a two-field, one-theme profile with a 0.5/0.5 field split and 9 distinct coauthors.
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.
- Identity is resolved from public bibliographic data via ORCID 0000-0002-1559-1048; institution and country are not recorded in the input data.
- Output, impact, and activity dimensions are unscored because publication-volume, citation, and multi-year cadence data are outside the current data scope.
- The profile rests on 2 Evolve-reviewed 2026 arXiv preprints; peer-review status is not recorded.
- Reproducibility (38) averages only 2 scored papers, and Evolve verdicts note missing statistical tests in both; the score may shift as new entries are added.