Zeming Liu

—

Unclaimed · AI-structured profileOpenAlex

Early-stage contributor to LLM agent memory and LLM evaluation statistics; two 2026 arXiv preprints

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-09-27) 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—

Insufficient data to score: the profile currently covers only two 2026 arXiv preprints, with no venue history or output-rate record. This axis would measure publication volume, venue quality, and productivity trend over time.

Impact—

Insufficient data to score: no citation counts, h-index, or adoption evidence is available for these preprints yet. This axis would measure citations, referencing works, and downstream influence on LLM agent memory and evaluation research.

Focus—

Insufficient data to score: with only two indexed papers, thematic concentration cannot be reliably computed, even though both lean toward LLM methodology. This axis would measure topical coherence across the publication record over successive years.

Collaboration44/100

Evolve scores this axis at 44 out of 100, computed from real bibliometric data: 8 distinct coauthors appear across the two cataloged 2026 preprints. Recurring collaborators include Yezhou Cheng, Runjia Du, Qibai Chen, Hang Lyu, Yankai Zeng, Bojun Lin, and Yilan Wei, indicating a compact, stable working group. With middle author positions and no corresponding authorship on either paper, the pattern reads as contributing team membership rather than consortium-level or lead-author collaboration.

Activity—

Insufficient data to score: the visible record spans a single year (2026) with no multi-year cadence. This axis would measure publication frequency, preprint posting rhythm, and sustained engagement across successive years.

Reproducibility—

Insufficient profile-level data: Evolve holds only per-paper readiness ratings (medium and high) for the two preprints, not aggregate artifact or replication evidence. This axis would measure code availability, artifact quality, and independent reproduction across the work.

Research summary

Zeming Liu is an early-stage researcher whose verifiable public record, as indexed in the Evolve catalog, currently comprises two arXiv preprints posted in 2026. The first, 'Scope Before You Persist: Preventing Cross-Family Interference in Agent Memory', concerns persistent memory for LLM-based agents: it proposes scoped retrieval to stop cross-family interference during continual adaptation and self-improvement, and Evolve's review credits it with solid paired evidence while cautioning that results come from one templated benchmark and a single model. The second, 'Sharp Limits for Honest Uncertainty in Hard-Budget Repeated Evaluation', is a theoretical contribution to LLM benchmarking, proving sharp replication limits with confidence intervals, minimax lower bounds, and stratified sampling under hard budget constraints, supported by matching designs and equal-budget evidence; Evolve rates its reproducibility readiness as high. On both papers Liu holds a middle author position and is not the corresponding author, suggesting a contributor rather than lead role so far. Collaboration metrics record 8 distinct coauthors and a collaboration score of 44 out of 100, consistent with a compact working-team pattern. No institution, country, or ORCID is present in the data, so the profile is name-resolved from public bibliographic records only.

Collaboration pattern

Across the two 2026 preprints, bibliometric data record 8 distinct coauthors and a collaboration score of 44/100. Recurring names — Yezhou Cheng, Runjia Du, Qibai Chen, Hang Lyu, Yankai Zeng, Bojun Lin, and Yilan Wei — point to a compact, recurring team. Liu sits in a middle author position on both papers without corresponding authorship, consistent with a contributing member rather than a project lead; no affiliation or cross-institution data is available to map partnerships further.

Research keywords

LLM agentspersistent memoryretrieval scopecross-family interferencecontinual adaptationrepeated evaluationconfidence intervalsminimax lower boundshard budget constraintsstratified samplingLLM benchmarkingstatistical learning theory

Highlights

Scoped retrieval to block cross-family interference in agent memory
In a 2026 arXiv preprint, Liu (middle author) contributes to a method that scopes persistent memory before persisting, preventing cross-family interference in LLM-agent self-improvement and continual adaptation; Evolve's review finds the paired evidence solid but limited to one templated benchmark and a single model, with medium reproducibility readiness.
Sharp minimax limits for honest uncertainty under hard evaluation budgets
A 2026 arXiv preprint on statistics for LLM evaluation proves sharp replication limits for repeated evaluation under hard budget constraints, using confidence intervals, minimax lower bounds, and stratified sampling, with matching designs and equal-budget evidence; Evolve rates its reproducibility readiness high, while noting that practical gains are regime-dependent.
A methodological throughline: measured claims about LLM systems
Taken together, the two 2026 preprints pair an engineering fix for agent memory reliability with a theory of evaluation uncertainty; both carry middle authorship, and Evolve's verdicts emphasize explicit scope caveats on each side — one templated benchmark and a single model for the memory work, regime-dependent practical gains for the theory.

Timeline

2026Co-authored 'Scope Before You Persist: Preventing Cross-Family Interference in Agent Memory' on arXiv in a middle author position.
2026Co-authored 'Sharp Limits for Honest Uncertainty in Hard-Budget Repeated Evaluation' on arXiv; Evolve rates its reproducibility readiness high.
2026Both preprints entered the Evolve review catalog, forming an early-stage profile centered on LLM agent memory and evaluation statistics, with 8 distinct coauthors recorded.

Papers in the Evolve catalog

Frequent co-authors

Yezhou Cheng×2 Runjia Du×2 Qibai Chen×2 Hang Lyu×2 Yankai Zeng×2 Bojun Lin×2 Yilan Wei×1 Yilan Wei×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 and collaboration metrics come from public bibliographic data (OpenAlex-style sources) only; the collaboration score of 44 and 8 distinct coauthors are the sole raw metrics provided.
No institution, country, or ORCID appears in the source data; 'Zeming Liu' is a common name, so this profile is name-resolved from public bibliographic records and conflation with other researchers is possible.
Coverage is limited to two 2026 arXiv preprints; no citation counts, venue history, or h-index are available, leaving five of six radar dimensions unscored.
Both papers record Liu in a middle author position and not as corresponding author; no further author-contribution details are available.