Alibaba researcher working across LLM agent reinforcement learning and multimodal retrieval embeddings
4
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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 data is insufficient: this axis would measure publication volume, venue quality, and first/corresponding-author share over time. Only four 2026 arXiv preprints, all middle-author, are in scope, so no score is assigned.
Impact—
Impact data is not yet available: this axis would track citations, h-index-style aggregates, and uptake of methods such as Rank-KL distillation. The four preprints are too recent (2026) for citation evidence, so no score is given.
Focus0/100
The focus score is 0, and the underlying field distribution is exactly 50/50 across two distinct lines — multimodal information retrieval (CORE) and LLM agent RL (ArenaFlow, ARISE-RL, MemoryWalker). The portfolio is deliberately dual-track rather than concentrated; the zero reflects split attention, not inactivity, and may converge as one line matures.
Collaboration71/100
Collaboration scores 71 on a 0-100 scale, with 34 distinct coauthors across the reviewed work and a recurring core of 12 frequent collaborators, including Ruixue Ding, Boli Chen, Zheng-Jun Zha, and Jieping Ye. This points to an active team-based researcher embedded in multi-author groups rather than a solo author, consistent with all four papers being middle-position, non-corresponding contributions.
Activity—
Activity data is insufficient: this axis would measure publication cadence, year-over-year output, and project turnover. All catalog entries fall in a single year (2026), so no longitudinal trend can be computed and no score is given.
Reproducibility55/100
Reproducibility scores 55 (mean over 2 numerically scored papers). Evolve's readiness labels rate ArenaFlow, CORE, and ARISE-RL as medium and MemoryWalker as low, with recurring blockers: judge circularity, missing variance tests, self-built benchmarks, and thin statistical rigor such as absent confidence intervals and seed reporting.
Research summary
Pengjun Xie is a researcher at Alibaba Group whose reviewed work splits evenly between two fast-moving areas of machine learning: multimodal information retrieval and LLM agent reinforcement learning (a 50/50 field distribution in the bibliometric snapshot). On the retrieval side, the CORE paper (2026) improves compositional reasoning in MLLM embeddings through Rank-KL reranker distillation, with controlled results though thin significance testing. On the agent side, three 2026 preprints address how agents learn and stay consistent: ArenaFlow extends tournament-based agent RL with step- and skill-level credit propagation, reporting +19.5 over ArenaRL on Open-Travel; ARISE-RL couples rubric-grounded rewards with generator-solver co-evolution and reward-gated self-distillation; MemoryWalker argues that training agents on compressed contexts they never saw breaks conditioning, proposing exact trajectory-tree walks plus SDCC. Across these reviews, Evolve consistently flags statistical rigor — judge circularity, missing variance tests, self-built benchmarks, thin confidence intervals — as the main caveat; reproducibility readiness is rated medium for three papers and low for one, with a mean numeric score of 55 over two scored papers. All four entries are 2026 arXiv preprints with middle-author, non-corresponding positions, so this remains an early-stage profile awaiting venue confirmation and citation evidence.
Collaboration pattern
Pengjun Xie publishes in multi-author teams: all four reviewed 2026 papers are middle-author, non-corresponding contributions, and the bibliometric snapshot counts 34 distinct coauthors with a stable core of 12 frequent collaborators, including Ruixue Ding, Boli Chen, Shihang Wang, Zheng-Jun Zha, and Jieping Ye. The network spans both research lines — multimodal information retrieval and LLM agent RL — consistent with project-based teaming rather than a fixed single-lab pattern.
Research keywords
Highlights
ArenaFlow: hierarchical credit propagation lifts open-ended agent RL
ARISE-RL: rubric-grounded self-evolution for open-ended agent RL
CORE: Rank-KL reranker distillation for compositional MLLM embeddings
MemoryWalker: training agents only on contexts they actually saw
Timeline
- 2026MemoryWalker — train-inference consistency for compressed-context LLM agent post-training (arXiv preprint, middle author).
- 2026ARISE-RL — rubric-grounded generator-solver co-evolution for open-ended agent RL (arXiv preprint, middle author).
- 2026CORE — Rank-KL reranker distillation improves compositional reasoning in MLLM embeddings (arXiv preprint, middle author).
- 2026ArenaFlow — trajectory ranking extended to hierarchical credit propagation in open-ended agent RL (arXiv preprint, middle author).
Papers in the Evolve catalog
ArenaFlow: From Trajectory Ranking to Hierarchical Credit Propagation for Open-Ended Agent RL
CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation
ARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning
MemoryWalker: Stop Training Agents on Contexts They Never Saw
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 indicators come from an OpenAlex-based snapshot only; publication volume, citation impact, and activity metrics are not yet available.
- The profile is name-resolved from public bibliographic data (ORCID 0009-0004-8412-359X); 'Pengjun Xie' may be ambiguous across researchers.
- All catalog entries are 2026 arXiv preprints; no peer-reviewed venue confirmations are included in this scope.
- The reproducibility score averages only 2 numerically scored papers; readiness labels for the other entries are qualitative.