Guanjun Jiang

Alibaba Group (Cayman Islands) · Cayman Islands

Unclaimed · AI-structured profileOpenAlex

Alibaba researcher building memory systems and reinforcement-learning methods for LLM agents

papers here

reviewed by Evolve

first-author

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—

The output axis would measure publication volume and productivity trajectory, but no score is computed yet: this OpenAlex-linked profile is too early-stage (h-index 1, i10-index 0) for a reliable aggregate.

Impact—

The impact axis would measure citation-based influence — citations, h-index derivatives and venue-weighted attention. No score exists yet: an h-index of 1 and i10-index of 0 on this early OpenAlex record are too thin to score honestly.

Focus0/100

Focus scores 0, the lowest of the four computed axes. The indexed field distribution splits evenly — machine learning/NLP for LLM agent memory, LLM agents/NLP, and LLM agents with reinforcement learning each take about one third — while indexed topics range widely (Topic Modeling 26, NLP Techniques 13, RL in Robotics 12, Domain Adaptation and Few-Shot Learning 9, Multimodal ML Applications 9).

Collaboration65/100

Collaboration scores 65, built on 25 distinct coauthors in the OpenAlex record. On the three Evolve-reviewed papers Jiang sits in the middle author position every time, never corresponding author — the signature of a contributor inside a stable industrial research team rather than a solo lead. A recurring core of coauthors (Fanyu Zhao, Liang Dong, Fugen Yao, Jian Xu, Han Zhang, among others) appears repeatedly, anchoring the team pattern.

Activity100/100

Activity scores 100, the maximum. The OpenAlex counts-by-year record shows one indexed work in 2025 jumping to 37 in 2026, with 38 works across the recent three-year window — a steep ramp-up concentrated in a single year. All three Evolve-reviewed papers carry 2026 arXiv dates, so the catalog and the bibliometric surge point to the same burst of current, full-time research output.

Reproducibility55/100

Reproducibility scores 55 (reproMean) across three reviewed papers, every one rated medium readiness. The pattern in Evolve's verdicts is consistent: methods and ablations are judged thorough — counterfactual credit in MemCalib, ablated gains in RPMem and Elastic Horizon — but evaluation caveats temper confidence, including LLM-judge circularity, multiple-choice-only testing, single-seed 14B runs, and margins close to seed noise.

Research summary

Guanjun Jiang is a researcher at Alibaba Group whose indexed work centers on large language model (LLM) agents — the memory systems agents use to persist and calibrate information, and the reinforcement-learning machinery used to train them. The Evolve-reviewed catalog holds three 2026 arXiv preprints, each with Jiang in a middle author position: MemCalib (a 15k-example benchmark for agent memory use paired with counterfactual-credit reinforcement learning), RPMem (recurrent parametric memory consolidated across sessions via hypernetworks and LoRA), and Elastic Horizon (a closed-loop controller that adapts the effective interaction horizon in agentic RL). Bibliometric signals are early-stage: the profile records an h-index of 1 and an i10-index of 0, so the output and impact axes remain unscored. Activity, by contrast, is extreme — one indexed work in 2025 versus 37 in 2026 (38 works across the recent three-year window), earning the maximum activity score of 100. Collaboration follows an industrial team pattern: 25 distinct coauthors, with a recurring core including Fanyu Zhao, Liang Dong, and Fugen Yao. Reproducibility scores 55 across the three reviewed papers, all rated medium readiness — Evolve credits thorough ablations while flagging caveats such as LLM-judge circularity and multiple-choice-only evaluation.

Collaboration pattern

Jiang's collaboration pattern is industrial-team research: 25 distinct coauthors in the OpenAlex record, and on all three Evolve-reviewed papers Jiang is a middle author and not the corresponding author. A tight recurring core — Fanyu Zhao, Liang Dong and Fugen Yao (repeated across coauthor lists), plus Jian Xu, Han Zhang, Yifei Zhao, Ruike Cao, Yinsheng Li and Li Xiao — suggests a stable Alibaba group working across agent memory, NLP and agentic reinforcement learning.

Research keywords

LLM agentsAgent memoryMemory calibrationParametric memoryLong-term memoryReinforcement learningAgentic reinforcement learningCredit assignmentBenchmarkHypernetworkLoRAInteraction horizon

Highlights

MemCalib: 15k-example benchmark and counterfactual-credit RL for agent memory
Introduces a benchmark for measuring and optimizing memory use in LLM agents, trained with counterfactual-credit reinforcement learning. Evolve judges it rigorous and well-ablated, but notes that LLM-judge circularity weakens the reliability of the conclusions. Reproducibility readiness: medium.
RPMem: recurrent parametric memory that persists across agent sessions
Proposes backbone-reusable recurrent parametric memory consolidated via hypernetworks and LoRA. Evolve reports strong, well-ablated gains on PERMA but modest improvements on other benchmarks, and flags that evaluation is multiple-choice only. Reproducibility readiness: medium.
Elastic Horizon: closed-loop control of the effective interaction horizon in agentic RL
Designs a horizon controller based on the P90 of success lengths with an exponential moving average, adapting the interaction frontier during agentic reinforcement learning. Evolve finds it well-ablated but notes 7B-scale margins near seed noise, a single-seed 14B result, and only two benchmarks. Reproducibility readiness: medium.

Timeline

2025First OpenAlex-indexed record appears — a single work in the counts-by-year data, before the current output surge.
2026MemCalib released on arXiv: benchmarking and optimizing memory use in LLM agents with counterfactual-credit reinforcement learning.
2026RPMem released on arXiv: learning long-term recurrent parametric memory across sessions for LLM agents.
2026Elastic Horizon released on arXiv: discovering the effective interaction frontier in agentic reinforcement learning.
2026OpenAlex records 37 indexed works in 2026 versus 1 in 2025 (38 works in the recent three-year window) — the steepest activity signal in this profile.

Papers in the Evolve catalog

Frequent co-authors

Fanyu Zhao×2 Liang Dong×2 Fugen Yao×2 Jian Xu×2 Han Zhang×2 Yifei Zhao×2 Ruike Cao×1 Yinsheng Li×1 Ruike Cao×1 Fugen Yao×1 Li Xiao×1 Fanyu Zhao×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.

Bibliometrics are computed from OpenAlex only; the identity is name-resolved from public bibliographic data (no ORCID), so common-name conflation cannot be fully ruled out.
Output and impact scores are missing because citation signals are too early-stage (h-index 1, i10-index 0); treat this as an early-stage profile.
The Evolve catalog covers only three 2026 arXiv preprints, all with middle authorship; verdicts reflect Evolve's reviews of those papers, not the full publication record.
No peer-reviewed venue, funding or award information is present in the provided data.