CAS researcher on LLM-agent memory calibration and graph-based retrieval-augmented generation for multi-hop QA
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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: this axis would measure publication volume and productivity over time, but no output metrics beyond the two reviewed arXiv papers are available in this snapshot, so no score is given and none is estimated.
Impact—
Insufficient data: this axis would measure citation-based influence such as citation counts, h-index, and field-weighted impact, but no citation statistics were provided for this researcher, so the axis remains unscored.
Focus0/100
The focus score is 0, derived from an exactly even field distribution: 50% machine learning (LLM agents / NLP) and 50% natural language processing (retrieval-augmented generation). The two catalogued papers split cleanly between agent-memory benchmarking and graph-based RAG, so the low score reflects breadth across two adjacent LLM subfields rather than scattered drift.
Collaboration61/100
The collaboration score is 61, based on 20 distinct coauthors appearing across the two catalogued papers. The frequent-coauthor list shows stable repeat partners — Fanyu Zhao, Liang Dong, and Guanjun Jiang each recur — alongside Ruike Cao, Fugen Yao, Yifei Zhao, Han Zhang, and Li Xiao. Xu occupies middle authorship positions rather than leading, consistent with team-based LLM research.
Activity—
Insufficient data: this axis would measure recent publication tempo and year-over-year consistency, but the reviewed sample only contains papers dated 2026, so no activity trend or score can be computed yet.
Reproducibility55/100
The reproducibility score of 55 is the mean over two rated papers (reproMean = 55, reproPapers = 2), both individually rated medium readiness. This aligns with Evolve's verdicts: MemCalib is rigorous and thoroughly ablated but leans on an LLM-judge loop, while Cognition on Graph reports large gains with thin statistics and high token cost — solid artifacts whose headline claims still need independent verification.
Research summary
Jian Xu is a researcher affiliated with the Chinese Academy of Sciences whose reviewed work sits at the intersection of LLM agents and retrieval-augmented generation. Evolve's catalog currently includes two 2026 arXiv papers with Xu in a middle authorship position, neither as corresponding author. The first, MemCalib, contributes a 15k-example benchmark for memory use in LLM agents together with a counterfactual-credit reinforcement-learning optimization method; Evolve judged it rigorous and thoroughly ablated, while flagging LLM-judge circularity as a limit on the strength of its conclusions. The second, Cognition on Graph, reports state-of-the-art multi-hop question answering through plan-explore-reflect cognitive cycles over Wikidata and Wikipedia, with large multi-metric gains but high token cost and thin statistical reporting. Bibliometric records show the reviewed output split exactly evenly (50/50) between machine learning (LLM agents/NLP) and natural language processing (retrieval-augmented generation), 20 distinct coauthors, and a mean reproducibility score of 55 across the two rated papers. This is an early-stage, small-sample profile: publication volume, citation impact, and activity trends are not yet available.
Collaboration pattern
Bibliographic records show 20 distinct coauthors across the two catalogued papers, yielding a collaboration score of 61. Repeat collaborators stand out: Fanyu Zhao, Liang Dong, and Guanjun Jiang each appear multiple times in the frequent-coauthor list, alongside Ruike Cao, Fugen Yao, Yifei Zhao, Han Zhang, and Li Xiao. On both reviewed papers Xu sits in a middle authorship position and is not corresponding author, pointing to a contributor role within team-based LLM research groups rather than solo or lead-senior work.
Research keywords
Highlights
15k-example benchmark for memory use in LLM agents
Counterfactual-credit RL for memory optimization, with a judge-circularity caveat
Plan-explore-reflect GraphRAG claiming SOTA multi-hop QA
Timeline
- 2026MemCalib posted to arXiv: a 15k-example memory-calibration benchmark for LLM agents, optimized with counterfactual-credit reinforcement learning.
- 2026Cognition on Graph posted to arXiv: cognitive cycles and bidirectional graph-text synergy over Wikidata and Wikipedia for multi-hop question answering.
- 2026Both papers currently reviewed by Evolve fall in 2026, making it the only year with reviewed output in this profile so far.
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.
- Jian Xu is a common name; this profile is name-resolved from public bibliographic data and the reviewed-paper match, with no ORCID provided for verification.
- Only two reviewed arXiv papers (both 2026, middle authorship, non-corresponding) underpin the qualitative assessment; publication counts, citations, and h-index are unavailable.
- Dimension scores are derived from the provided bibliometric snapshot; underlying database coverage and time windows are not specified, and output, impact, and activity are unscored due to missing data.
- Reproducibility figures (mean 55) rest on only two rated papers, so the score should be read as indicative rather than stable.