RL for LLM agents and retrieval: curriculum-rewarded tool learning and reasoning-augmented reranking.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
Not yet scored. This axis would measure publication volume and productivity over time — total papers, annual output, and venue spread — but the current snapshot contains only the two reviewed arXiv preprints, too thin to compute.
Impact—
Not yet scored. Impact would track citations, downstream usage and influence of the reviewed papers, but no citation counts or venue acceptance data are available for these 2026 preprints, so no estimate is made here.
Focus0/100
The focus score is 0, reflecting a perfectly even field split: the raw distribution assigns 0.5 to machine learning (LLM tool learning / reinforcement learning for agents) and 0.5 to information retrieval / natural language processing. Both cataloged 2026 papers sit one per field, yet they share a GRPO-based reinforcement-learning methodology, so the low concentration reads as an early-stage, two-front agenda rather than scattered work.
Collaboration48/100
The collaboration score is 48, based on 10 distinct coauthors across the two reviewed papers. The frequent-coauthor list is headed by Zhengzong Chen, Yuanyuan Zhao and Fei Huang, who recur most often, alongside Shihao Liu, Hao Yin and Wenyan Li. Authorship position varies — first author on the reranking paper, middle author on the tool-learning paper — and no corresponding-author role appears in the current record.
Activity—
Not yet scored. Activity would capture publishing cadence and recency — papers per year, gaps between releases, and ongoing work. The snapshot holds only two 2026 arXiv preprints, too few to characterize a productive rhythm.
Reproducibility38/100
Reproducibility scores 38, the mean readiness across the 2 assessed papers. The tool-learning paper (MATCH) rates medium: Evolve notes strong ablations but single-run results without variance tests. The reranking paper (MERIT-Rank) rates low, with efficiency claims and reproducibility explicitly flagged for scrutiny. The pattern suggests careful experimental design undermined by limited statistical reporting.
Research summary
Lijun Liu is an early-stage researcher whose cataloged work spans two complementary strands of large language model research: reinforcement-learning-based tool learning for LLM agents, and reasoning-augmented text reranking in information retrieval. The field distribution splits exactly evenly — 0.5 machine learning (LLM tool learning / reinforcement learning for agents), 0.5 information retrieval / natural language processing — producing a focus score of 0 and signaling a two-front agenda rather than a settled specialization. A clear methodological thread connects both 2026 arXiv preprints: GRPO-based reinforcement learning. MATCH couples curriculum scheduling with hierarchically gated rewards for tool learning, while MERIT-Rank, the first-author paper, layers multi-perspective evidence and reasoning into staged reranking evaluated on BRIGHT, TREC and BEIR. Collaboration is moderate — 10 distinct coauthors and a collaboration score of 48 — anchored by recurring partners Zhengzong Chen, Yuanyuan Zhao and Fei Huang, with authorship ranging from first to middle position and no corresponding-author record. Reproducibility is the weakest documented axis: a readiness mean of 38 across 2 papers, with Evolve flagging single-run results lacking variance tests and efficiency claims needing scrutiny. Output, impact and activity metrics are not yet available in the underlying data.
Collaboration pattern
The collaboration footprint is moderate: a collaboration score of 48 with 10 distinct coauthors across two papers. Zhengzong Chen, Yuanyuan Zhao and Fei Huang recur most often in the provided co-author list, with Shihao Liu, Hao Yin and Wenyan Li also present. Authorship varies between first author (reranking) and middle author (tool learning), with no corresponding-author credit recorded. The recurring core suggests a stable small team spanning both the tool-learning and retrieval lines.
Research keywords
Highlights
MATCH: curriculum-scheduled RL for LLM tool learning
MERIT-Rank: multi-perspective reasoning for text reranking
Timeline
- 2026MATCH, a model-aware tool-learning paper combining curriculum scheduling and hierarchically gated rewards, is cataloged from arXiv with middle authorship.
- 2026Think Thrice Before Reranking (MERIT-Rank) appears on arXiv as a first-author paper on multi-perspective evidence and reasoning integration for text reranking.
- 2026Both preprints enter the Evolve review catalog, yielding a reproducibility readiness mean of 38 across 2 papers and a 10-distinct-coauthor collaboration record.
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.
- Identity is name-resolved from public bibliographic data; 'Lijun Liu' is a common name and no ORCID, institution or country is attached to this profile.
- Radar scores are computed from public bibliometric data only; output, impact and activity axes lack underlying data and are left unscored.
- Both cataloged papers are 2026 arXiv preprints; no peer-reviewed venue records are included in the current snapshot.
- Reproducibility (mean 38) reflects Evolve's review of only 2 papers and may not generalize beyond this small sample.