Alibaba Group researcher on LLM agent memory: calibration benchmarks and recurrent parametric long-term memory.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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—
The output axis is unscored because publication-count data is insufficient in the current snapshot. This axis would measure research productivity — the number of papers over time, by venue and authorship role — and will populate as more bibliographic records are indexed.
Impact—
The impact axis is unscored because citation-based data is not yet available. It would measure scholarly influence — citations, referencing works, and downstream adoption — and cannot be estimated from the two cataloged arXiv papers alone.
Focus0/100
The radar focus axis is scored 0, which contrasts with the field distribution across the two cataloged papers: 0.5 for machine learning / natural language processing (LLM agent memory) and 0.5 for machine learning (LLM agents / NLP). Both labels point to the same narrow topic — memory in LLM agents — so topical concentration appears high even though the score records 0; the gap likely reflects how few papers feed the metric.
Collaboration58/100
The collaboration axis scores 58, based on 17 distinct coauthors across the cataloged record. The most frequently recurring names are Fanyu Zhao, Fugen Yao, and Ruike Cao (each appearing twice in the frequent-coauthor list), alongside Jian Xu, Guanjun Jiang, Han Zhang, Yifei Zhao, Yinsheng Li, and Li Xiao. On both 2026 papers Liang Dong holds a middle author position without corresponding status, indicating a contributing role within a stable Alibaba-affiliated group rather than sole leadership.
Activity—
The activity axis is unscored because temporal publication data is insufficient. It would measure research cadence — papers per year, recency, and continuity of output — and needs a longer indexed publication window than the current record provides.
Reproducibility55/100
The reproducibility axis scores 55, a mean computed over 2 papers with recorded readiness; both MemCalib and RPMem are rated medium reproducibility readiness. Evolve's reviews note rigorous, well-ablated experiments in each paper, but flag evaluation caveats — LLM-judge circularity in MemCalib and multiple-choice-only evaluation in RPMem — which temper how confidently an independent team could reproduce the reported results end-to-end.
Research summary
Liang Dong is a researcher listed with Alibaba Group (Cayman Islands) whose cataloged work centers on memory in LLM agents. The Evolve catalog contains two 2026 arXiv papers, on both of which Liang Dong appears as a middle author without corresponding status. The first, MemCalib, introduces a 15,000-example benchmark for calibrating memory use in LLM agents together with a counterfactual-credit reinforcement-learning optimization method; Evolve's review judges it rigorous and well-ablated while cautioning that LLM-judge circularity weakens the conclusions. The second, RPMem, proposes a recurrent parametric memory consolidated across sessions via hypernetwork and LoRA components, claimed to be reusable across backbones; it shows strong, well-ablated gains on PERMA but modest effects elsewhere, with evaluation limited to multiple-choice formats. The field distribution splits evenly (0.5 / 0.5) between two closely related labels, both pointing to LLM-agent memory, indicating a tightly focused footprint across the current catalog sample. The collaboration axis scores 58 from 17 distinct coauthors, with Fanyu Zhao, Fugen Yao, and Ruike Cao recurring most often in the record. Reproducibility scores 55 as a mean over the two papers, both rated medium readiness. Output, impact, and activity axes lack sufficient data and remain unscored; citation counts, h-index, venue history, and ORCID are not available in the current snapshot, so this profile should be read as an early-stage, name-resolved record.
Collaboration pattern
Liang Dong's cataloged work involves 17 distinct coauthors, and the collaboration axis scores 58. The most frequently recurring names are Fanyu Zhao, Fugen Yao, and Ruike Cao (each listed twice among frequent coauthors), together with Jian Xu, Guanjun Jiang, Han Zhang, Yifei Zhao, Yinsheng Li, and Li Xiao. On both 2026 papers Liang Dong is a middle author without corresponding status, consistent with a contributing specialist role inside a stable Alibaba-affiliated team.
Research keywords
Highlights
MemCalib: benchmarking and optimizing memory use in LLM agents
RPMem: recurrent parametric long-term memory across sessions
Timeline
- 2026MemCalib posted to arXiv: a benchmark plus counterfactual-credit reinforcement learning for memory use in LLM agents.
- 2026RPMem posted to arXiv: recurrent parametric memory consolidated across sessions for LLM agents.
- 2026Both papers entered the Evolve review catalog, each rated medium reproducibility readiness.
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.
- Bibliometric radar scores are computed from public bibliometric data (OpenAlex-derived); the output, impact, and activity axes are unscored due to insufficient data.
- "Liang Dong" is a common name; this profile is name-resolved from public bibliographic data and may merge or split identities.
- Only two papers (both 2026 arXiv preprints) are in the Evolve catalog; citation counts, h-index, venue history, and ORCID are not available.
- The focus axis records 0 even though the field distribution concentrates on LLM-agent memory; the low score reflects the sparse sample and should be read with caution.