Alibaba researcher advancing LLM agent memory and distillation: MemCalib benchmark, RPMem, on-policy distillation.
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
Publication output metrics — total paper count, venue mix, and first/corresponding-author share — are not yet available; the catalog holds only three 2026 arXiv entries, so this axis cannot be scored yet.
Impact—
Impact data such as citation counts, h-index, and citing works are absent from the current snapshot; this axis would measure how widely the three reviewed papers are cited and adopted by other research.
Focus0/100
Focus scores 0, reflecting a deliberately spread portfolio: the field distribution splits evenly at one third each across LLM training and distillation, LLM agent memory, and LLM agents/NLP. The three lines nonetheless converge on one theme — making LLM agents reliable — so the breadth reads as thematic extension rather than true dispersion.
Collaboration65/100
Collaboration scores 65, based on 25 distinct coauthors, 12 of whom recur across the reviewed papers (including Fanyu Zhao, Liang Dong, and Han Zhang). On all three 2026 papers Jian Xu sits in a middle authorship position without a corresponding-author role, suggesting a steady contributor within a larger Alibaba group effort rather than a solo lead.
Activity—
Activity signals — yearly publication counts and recent output trend — are insufficient, since only three 2026 papers are on record; productivity pace and sustained research cadence cannot be assessed yet.
Reproducibility55/100
Reproducibility scores 55 (mean across 3 papers with readiness data), and every reviewed paper is rated medium. Evolve's verdicts flag concrete limits: code unreleased for the on-policy distillation work, LLM-judge circularity weakening MemCalib's conclusions, and multiple-choice-only evaluation for RPMem. Method sections are described as well-ablated, so reproduction is feasible but needs re-implementation effort.
Research summary
Jian Xu is a researcher at Alibaba Group whose reviewed work in Evolve's catalog centers on the reliability of large language model (LLM) agents and the training methods behind them. The 2026 arXiv record spans three tightly linked lines. MemCalib introduces a benchmark — described in Evolve's verdict as roughly 15,000 examples — for measuring and optimizing how agents use memory, paired with counterfactual-credit reinforcement learning. RPMem proposes a recurrent parametric memory, built with hypernetworks and LoRA, that consolidates experience across sessions and reports strong gains on PERMA. 'Teacher Should Think Ahead' adaptively extends teacher continuation length in on-policy distillation, supported by a variance-bias analysis and solid Qwen3 math and code gains. On all three papers Xu holds a middle authorship position without a corresponding role, consistent with contributing inside a larger Alibaba team alongside recurring collaborators such as Fanyu Zhao, Liang Dong, and Han Zhang. Evolve consistently credits rigorous ablations and solid empirical results while flagging reproducibility caveats: unreleased code, LLM-judge circularity in MemCalib, and multiple-choice-only evaluation for RPMem. Because 'Jian Xu' is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Jian Xu works within a sizable Alibaba collaboration network: 25 distinct coauthors, 12 of whom recur across the reviewed papers — including Fanyu Zhao, Liang Dong, Fugen Yao, Guanjun Jiang, and Han Zhang. On all three 2026 papers the position is middle-author without a corresponding role, suggesting a contributing specialist inside team-led projects rather than a solo group lead.
Research keywords
Highlights
MemCalib: benchmarking and optimizing memory use in LLM agents
RPMem: recurrent parametric memory for cross-session LLM agents
Adaptive teacher continuations for reliable on-policy distillation
Timeline
- 2026Co-authored MemCalib, a memory-calibration benchmark (~15,000 examples per Evolve's verdict) for LLM agent memory use, combined with counterfactual-credit reinforcement learning.
- 2026Co-authored RPMem, introducing recurrent parametric memory built with hypernetworks and LoRA for cross-session LLM agent memory, with strong gains on PERMA.
- 2026Co-authored 'Teacher Should Think Ahead', an adaptive teacher-continuation method for on-policy distillation with variance-bias theory and Qwen3 math/code gains.
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.
- Radar scores and bibliometrics are derived from OpenAlex-based public data; the output, impact, and activity axes lack sufficient data and are left unscored.
- Coverage is limited to three 2026 arXiv papers reviewed by Evolve; earlier or non-arXiv work by this author may be missing.
- 'Jian Xu' is a common name; this profile is name-resolved from public bibliographic data, with affiliation recorded as Alibaba Group (Cayman Islands) and no country-level detail.
- All three papers carry only a medium reproducibility rating; code release is explicitly noted as missing for the distillation paper.