Zirui Li

Hong Kong Polytechnic University

Unclaimed · AI-structured profileOpenAlex

Evidence-grounded NLP researcher (Hong Kong PolyU) in citation faithfulness verification and clinical mental-health NLP

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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 is not yet scored: the catalog holds only two 2026 arXiv entries, too few to estimate publication volume or authorship trajectory. This axis would measure total outputs, venue quality, and author-role distribution across years.

Impact—

Impact cannot be scored yet: no citation counts, h-index, or downstream-uptake data were provided for the two 2026 papers. This axis would measure citations, uptake of the methods, and influence within the NLP community.

Focus0/100

Focus scores 0/100, reflecting an even 50/50 split between natural language processing and clinical NLP for computational mental health across the two catalog papers. One paper verifies citation faithfulness with DeBERTa ensembles; the first-authored paper targets suicide risk detection from LLM internal representations. The portfolio spans two adjacent but distinct problem spaces rather than a single concentrated line.

Collaboration28/100

Collaboration scores 28/100 with only 3 distinct coauthors recorded across the two papers: Yanling Li, Kaolanglang Gao, and Mingyu Wan. The first-authored suicide-risk paper suggests independent leadership, while the NLPCC Task 10 entry was a middle-position team effort. The low distinct-coauthor count and score indicate a small, stable working circle rather than a broad collaborative network.

Activity—

Activity is not yet scored: both catalog entries are dated 2026, so year-over-year publishing cadence cannot be computed. This axis would measure publication frequency, recent output momentum, and continuity of research activity over time.

Reproducibility55/100

Reproducibility scores 55/100, a mean computed over 2 papers, both currently rated medium on reproducibility readiness. The NLPCC Task 10 paper earned reviewer credit for thorough ablations and honest limitations, though its key conclusions rest on out-of-fold diagnostics without significance testing. The suicide-risk detection paper reports small but confirmation-validated gains, suggesting careful but not yet fully turnkey empirical practices.

Research summary

Zirui Li is an early-career NLP researcher at Hong Kong Polytechnic University whose catalog splits evenly — 50/50 in the recorded field distribution — between natural language processing and clinical NLP for computational mental health. The catalog contains two 2026 arXiv papers. The first-authored study, 'Beyond Final-Token Classification,' proposes heterogeneous readouts (HRD) that replace final-token margins with latent readouts of LLM internal representations for evidence-grounded suicide risk detection, using QLoRA fine-tuning and probing classifiers to deliver small but confirmation-validated gains. The second paper, a middle-position contribution to NLPCC 2026 Task 10, verifies citation-level faithfulness with DeBERTa ensembles and class-wise calibration, finishing as Track 2 runner-up at 82.99 overall and earning reviewer credit for thorough ablations and honest limitations. Collaboration is currently compact: 3 distinct coauthors — Yanling Li, Kaolanglang Gao, and Mingyu Wan — with a collaboration score of 28/100, consistent with a small stable team. Reproducibility scores 55/100 across 2 papers rated medium on readiness, supported by careful diagnostics. Output, impact, and activity are not yet scored because the available bibliometric base is too thin. This is an early-stage profile built from a small evidence set.

Collaboration pattern

The catalog records 3 distinct coauthors — Yanling Li, Kaolanglang Gao, and Mingyu Wan — yielding a low collaboration score of 28/100. The pattern suggests a compact, stable working group: independent first-author leadership on the suicide-risk detection study, and a middle-position role within the NLPCC Task 10 team. No cross-institution or senior-collaborator links are visible in the provided data, so the network remains small and locally concentrated.

Research keywords

scientific claim verificationcitation faithfulnessevidence retrievalDeBERTaensemble learningclass-wise calibrationsuicide risk detectionLLM internal representationsprobing classifiersQLoRA fine-tuningmulti-label classificationevidence extraction

Highlights

NLPCC 2026 Track 2 Runner-Up in Citation Faithfulness Verification
As a middle-position author, contributed to a DeBERTa ensemble system with class-wise calibration that scored 82.99 overall in NLPCC 2026 Task 10; the review noted thorough ablations and honest limitations, while key conclusions rest on out-of-fold diagnostics without significance tests.
First-Author Method for Evidence-Grounded Suicide Risk Detection
Led the heterogeneous readouts (HRD) study, which replaces final-token classification margins with latent readouts of LLM internal representations — using QLoRA fine-tuning and probing classifiers for multi-label suicide risk detection — delivering small but confirmation-validated gains.
A Consistent Evidence-Centric Research Thread
Both 2026 papers are organized around evidence — evidence retrieval and calibration on one side, evidence extraction from clinical text on the other — pointing to an emerging agenda that links verification methods with computational mental health applications.

Timeline

2026Posted first-authored work on heterogeneous readouts for evidence-grounded suicide risk detection to arXiv, entering clinical NLP for computational mental health.
2026Contributed as a middle-position author to the NLPCC 2026 Task 10 entry on citation-level faithfulness verification; the DeBERTa ensemble system finished as Track 2 runner-up with 82.99 overall.
2026Both papers entered the Evolve review catalog with medium reproducibility readiness, forming an early-stage dual-field profile split 50/50 between NLP and clinical NLP.

Papers in the Evolve catalog

Frequent co-authors

Yanling Li×2 Kaolanglang Gao×1 Mingyu Wan×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.

All bibliometric scores rest on a very small base: 2 catalog papers and 3 distinct coauthors; treat every dimension as an early-stage signal.
No citation counts, h-index, or publication-volume data were available, so output, impact, and activity remain unscored.
No ORCID was provided and Zirui Li is a common name; this profile is name-resolved from public bibliographic data and may merge or split identities.
Both papers are 2026 arXiv entries; peer-reviewed venue acceptance beyond the NLPCC shared-task runner-up result is not recorded.