Yingrui Li

—

Unclaimed · AI-structured profileORCID OpenAlex

NLP researcher dissecting evaluation scores: fact-verification evidence accounting and controlled long-context QA

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-02) 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—

No output metrics are available yet. This axis would measure cataloged publication volume, venue types, and cumulative output over time, but the underlying record is too sparse to compute them; it is left unscored rather than estimated.

Impact—

Impact data — citations, h-index, and downstream uptake — is not yet available for this record. The axis would measure how widely the two cataloged 2026 papers are cited and built upon; it remains unscored rather than guessed.

Focus0/100

The focus dimension scores 0, reflecting minimal concentration in the cataloged record: field distribution splits evenly (0.5/0.5) between fact-verification evaluation methodology and NLP/information retrieval for long-context QA. The two strands are adjacent — both ask how retrieval-and-verification pipelines should be evaluated — but with only two 2026 papers, this breadth may reflect a small record rather than settled orientation.

Collaboration14/100

The collaboration dimension scores 14, consistent with a very small network: the record lists exactly 1 distinct coauthor, Han Chen, who appears across both 2026 papers. Li takes first authorship on one paper and last authorship on the other, though neither lists Li as corresponding author. The pattern suggests a stable two-person working unit; with only two cataloged papers, broader network effects cannot yet be assessed.

Activity—

Activity metrics such as publication cadence and recent output rate lack the underlying data needed for scoring. With only two 2026 papers indexed, yearly production trends cannot be computed; the axis stays unscored until a fuller record exists.

Reproducibility55/100

Reproducibility averages 55 across the 2 papers Evolve has rated (reproPapers = 2), with both assigned medium readiness. This mid-range score matches the reviews: controlled designs, exact 2×2 recombination, and rigorous statistics support replication, while scope limits — DCUF-trained verifiers, a weak planner, LongBench-v2 only — temper confidence in generalizing the findings.

Research summary

Yingrui Li (ORCID 0000-0001-6630-1277) is an early-stage natural language processing researcher whose cataloged work centers on evaluation methodology for fact verification and retrieval-based question answering. The Evolve catalog holds two arXiv papers, both dated 2026. As first author, Li produced a controlled negative-result study of planning, routing, and reranking for long-context QA: SFT-trained evidence planners failed to beat strong hybrid retrieval on LongBench-v2, a result the review credits for rigorous controls while noting a narrow scope and a weak planner. As last author, Li contributed an exact 2×2 recombination across trained verifiers and LLMs showing that most of a 9.61-point joint fact-verification gain traces to evidence rather than answer accounting, with mixed LLM effects and scope limited to DCUF-trained verifiers. A consistent signature runs through both papers: controlled designs, careful statistics, and honest reporting of negative or scope-limited findings. Reproducibility averages 55 across the two rated papers, each rated medium. Collaboration is compact — one distinct coauthor, Han Chen. Institutional affiliation and country are absent from the record, so this profile is name-resolved from public bibliographic data.

Collaboration pattern

The cataloged collaboration pattern is compact: exactly 1 distinct coauthor, Han Chen, recurs across both 2026 papers, consistent with a stable two-person team. Li is first author on the long-context QA study and last author on the fact-verification decomposition, but is not listed as corresponding author on either. No institutional affiliation appears in the record, so lab ties and advisor relationships cannot be assessed from available data.

Research keywords

fact verificationFEVEROUSevidence retrievalscore decompositionShapley attributionDeBERTalong-context QAretrieval-augmented generationlearned context planningevidence selectionLongBench-v2hybrid retrieval

Highlights

Most of a 9.61-point fact-verification gain is evidence, not answers
An exact 2×2 recombination across trained verifiers and LLMs decomposes the joint improvement: the bulk traces to evidence quality rather than answer accounting, with rigorous statistics but scope limited to DCUF-trained verifiers and mixed LLM effects.
Learned context planning fails to beat strong retrieval in a controlled test
This first-author study isolates planning, routing, and reranking for long-context QA on LongBench-v2: SFT-trained evidence planners lose to strong hybrid retrieval, a negative result delivered with rigorous controls, though scope is narrow and the planner weak.
A consistent signature: controlled designs and honest negative results
Across both 2026 papers the same recipe recurs — exact recombination, careful controls, explicit scope limits — earning a reproducibility mean of 55 across two medium-readiness ratings in the Evolve catalog.

Timeline

2026Publishes a first-author controlled study on arXiv showing learned context planning fails to beat strong retrieval for long-context QA on LongBench-v2.
2026Publishes a last-author arXiv paper applying exact 2×2 recombination to fact verification, showing most of a 9.61-point joint gain is evidence-driven.
2026Both papers enter the Evolve catalog and receive medium reproducibility ratings, establishing a mean of 55 with one distinct coauthor on record.

Papers in the Evolve catalog

Frequent co-authors

Han Chen×2

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.

This is a common name; the profile is resolved from public bibliographic data tied to ORCID 0000-0001-6630-1277, with no institutional affiliation or country in the record.
Bibliometric dimensions are computed from a small snapshot covering only the 2 papers in the Evolve catalog; output, impact, and activity axes lack data and are unscored.
Both cataloged papers are 2026 arXiv preprints; peer-review status and publication venues are unknown.
Collaboration (14) and focus (0) scores rest on two papers and one coauthor and may shift substantially as the record grows.