NLP researcher at Peng Cheng Laboratory spanning LLM evaluation, agent memory, and efficient LLM architectures
3
papers here
3
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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 metrics are not yet available: this OpenAlex record counts only 1 work in 2026 and no cumulative publication totals, so no score is given. This axis would measure publication volume, venue spread, and output trajectory over time.
Impact—
Citation-impact data is insufficient: the record shows h-index 0 and i10-index 0 with no citation history, which most plausibly reflects an unresolved common-name profile rather than measured impact. This axis would measure citations, h-index, and influence on follow-up work.
Focus0/100
The focus score is 0, and the field distribution supports that reading: the record splits evenly across natural language processing (LLM evaluation), NLP (conversational AI and agent memory), and efficient LLM architecture, each at 0.333, while the three indexed topics each carry a single work. With only one recent work, no direction yet dominates — an even three-way spread rather than demonstrated specialization.
Collaboration86/100
The collaboration score is 86, the strongest axis in the record: OpenAlex indexes 73 distinct coauthors, and 12 recur often enough to be flagged, including Zihan Gu, Zhiyuan Wang, and Tianyi Ma. The three 2026 catalog papers place this researcher in different team settings — first-authoring a benchmark while holding middle positions on a large-model systems paper and a conversational-memory paper — indicating broad, cross-team co-authorship rather than a single fixed group.
Activity20/100
The activity score is 20, matching a thin recent record: countsByYear lists a single work in 2026 and recent3yWorks is 1, so there is no visible multi-year publication cadence. All three Evolve-cataloged papers carry 2026 arXiv dates, suggesting the visible activity is clustered in one year, but the underlying OpenAlex profile likely undercounts — a caveat rather than a confirmed productivity gap.
Reproducibility32/100
The reproducibility score is 32, a reproMean computed over 3 reviewed papers. Evolve rates the first-author Neo-Classic benchmark at medium readiness but flags weak statistics; the KV-cache-compression paper is self-evaluated with no ablations and mixed Table 1 results; the memory-RL paper shows strong gains yet LLM-judge circularity. Both middle-author papers carry low readiness, so the average sits in the lower third.
Research summary
Han Zhang is a researcher at Peng Cheng Laboratory (China) whose visible record spans three evenly weighted directions: natural language processing for LLM evaluation, conversational AI with agent memory, and efficient large language model architecture — each at 0.333 in the field distribution of an OpenAlex-resolved profile (no ORCID, so identity is name-resolved from public bibliographic data). The cataloged output is concentrated in 2026: three arXiv papers, first-authored on one. Neo-Classic (2026) is a first-author benchmark of 1,406 contemporary metrical classical Chinese poems that measures a 20–50% memorization gap and near-zero reordering behavior in LLMs; Evolve judges the design credible but the statistics weak. Two middle-author papers extend into systems and agents: DeepSeek-V4.1-Flash (2026) reports a 552B-parameter MoE claiming frontier-agent performance at 1/4–1/8 of the KV cache, and Interactive Memory Learning (2026) trains two-agent memory with reinforcement learning under delayed rewards. Bibliometric signals are thin — h-index 0, i10-index 0, 1 work in the 2026 count, and 1 work over the recent three years — which most plausibly reflects an incomplete common-name record rather than established impact. Collaboration is the standout signal, with 73 distinct coauthors and a collaboration score of 86; reproducibility averages 32 across the three reviewed papers.
Collaboration pattern
Collaboration is the strongest signal in this record: 73 distinct coauthors are indexed and 12 recur — Zihan Gu, Zhiyuan Wang, Tianyi Ma, Jiacheng Lu, Xinyan Zhang, Yuhao Wei, Cheng Hua, Cai Ke, Jiangyue Yan, Xin Liu, Zike Yuan, and Yue Yu. The pattern spans team types: an independent first-author benchmark alongside middle-author roles on a large systems paper and a conversational-agent paper, consistent with cross-team work rather than one fixed group.
Research keywords
Highlights
First-author benchmark exposing memorization in classical Chinese poetry
Co-author on KV-cache compression at frontier scale
Reinforcement learning for long-term conversational memory
Timeline
- 2026First-authored Neo-Classic, a benchmark of 1,406 metrical poems probing linguistic-aesthetic reasoning in classical Chinese poetry (arXiv).
- 2026Co-authored (middle position) DeepSeek-V4.1-Flash on KV cache compression for a 552B-parameter MoE (arXiv).
- 2026Co-authored (middle position) Interactive Memory Learning, applying two-agent reinforcement learning to long-term conversation memory (arXiv).
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.
- Name-resolved from a single OpenAlex record without ORCID; 'Han Zhang' is a common name, so h-index 0 and i10-index 0 likely reflect incomplete disambiguation rather than measured impact.
- Bibliometrics are OpenAlex-only: the output and impact axes are unscored because the record carries just 1 work in 2026 and no citation history.
- All three reviewed papers are 2026 arXiv preprints; no peer-reviewed venue, funding, or award data is in scope.
- The reproducibility score of 32 reflects Evolve's review of only 3 papers; co-author statistics describe volume, not collaboration quality.