Lei Zhang

—

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher spanning optimal transport theory, multimodal LLMs, and RL agent reliability

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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—

No publication-count, venue-tier, or authorship-share data is available in this snapshot, so output is unscored. This axis would measure total research productivity and its distribution across venues, years, and author positions.

Impact—

Citation counts, h-index, and downstream uptake are absent from the current data, so impact is unscored. This axis would measure how often and how widely the community cites and builds on this researcher's work.

Focus0/100

Within the reviewed catalog, output splits exactly evenly across three fields — optimal transport with computational biology applications, multimodal vision-language models, and NLP/LLM agents — at 33.3% each, producing a focus score of 0/100. This is a descriptive signal of breadth, not a quality judgment: each line has its own methods, benchmarks, and collaborator clusters.

Collaboration62/100

The three reviewed papers involve 16 distinct coauthors, with a recurring core of 12 named collaborators — including Junda Ying, Zhiwei Zeng, and Peijie Zhou — yielding a collaboration score of 56/100. Zhang's positions (last author on the optimal transport paper, middle author on the other two) suggest senior or supervisory roles on theory work and technical contributions within larger applied teams.

Activity—

Beyond the three 2026 arXiv preprints, no multi-year publication cadence is available, so activity is unscored. This axis would measure publishing pace and sustained research engagement across recent years.

Reproducibility55/100

Reproducibility readiness averages 55/100 over the three reviewed papers, each individually rated medium. Consistent with that, the platform's verdicts flag modest or untested empirical gains for the optimal transport and AREA results, and single-seed, small-scale evidence for the necessity-reward fix in the RL tool-use study — all worth independent verification before reuse.

Research summary

Lei Zhang is a machine learning researcher whose reviewed work spans three areas in near-equal measure: optimal transport with computational biology applications, multimodal vision-language modeling, and natural language processing for LLM agents. The Evolve catalog holds three 2026 arXiv preprints. The most theory-heavy, Dynamic Generalized Gromov-Wasserstein Optimal Transport, places Zhang in the last-author position; it develops TP-DATE, combining dynamic quadratic-form optimal transport theory with a simulation-free travelling-pair flow-matching solver aimed at applications including spatial transcriptomics, and reviewers judged the theory solid and the empirical gains real but modest. In the multimodal line, Zhang contributed to AREA (Layers, Sinks, and Scaling), a training-free inference-time method that adaptively gates, scales, and refreshes evidence highlighting, beating LoT on all nine knowledge-based VQA settings. The third paper, Spurious Tool Use, is a controlled study showing reinforcement-learning agents can learn cue-driven spurious tool use (up to +39.2%), corrected by a judge-based necessity reward. Bibliometric signals show moderate collaboration breadth (16 distinct coauthors; collaboration 56/100) and medium reproducibility readiness (mean 55/100 across three papers). The even 33.3/33.3/33.3 field split yields a focus score of 0/100 — an early-stage, deliberately broad portfolio. Output, impact, and activity axes lack underlying data and remain unscored.

Collaboration pattern

Zhang's reviewed work is team-based: 16 distinct coauthors across three papers, with a recurring group of 12 collaborators (including Junda Ying, Zhiwei Zeng, Peijie Zhou, Zhenbin Wang, Lituan Wang, and Wei Huang) suggesting stable working clusters. The collaboration score is 56/100. Authorship positions — last author on the optimal transport theory paper and middle author on the two LLM-oriented papers — point to a mix of senior theoretical leadership and collaborative contribution within larger groups.

Research keywords

Optimal transportGromov-Wasserstein distanceFlow matchingSimulation-free solversSpatial transcriptomicsMultimodal large language modelsKnowledge-based visual question answeringRetrieval-augmented generationReinforcement learningLLM agents and tool useSpurious correlation and shortcut learning

Highlights

Dynamic Gromov-Wasserstein optimal transport with simulation-free flow matching
Last-author paper introducing TP-DATE, which extends dynamic quadratic-form optimal transport theory and adds a simulation-free travelling-pair flow-matching solver, with spatial transcriptomics among the application targets; Evolve rates the theory solid and the empirical gains real but modest, with significance untested.
Training-free AREA evidence selection for multimodal LLMs
Contributed to AREA (Layers, Sinks, and Scaling), a training-free inference-time intervention that adaptively gates, scales, and refreshes evidence highlighting for knowledge-based VQA, beating LoT on all nine settings; gains are modest and untested, with medium reproducibility readiness.
Diagnosing spurious tool use in RL agents
Controlled study showing reinforcement-learning agents can learn cue-driven spurious tool use — lifting success by up to +39.2% — and that a judge-based necessity reward corrects it; the evidence remains single-seed and small-scale.

Timeline

2026Last-author preprint on dynamic generalized Gromov-Wasserstein optimal transport (TP-DATE), extending dynamic quadratic-form OT with simulation-free travelling-pair flow matching, targeting applications including spatial transcriptomics.
2026Co-authored the AREA paper on adaptive training-free evidence selection for multimodal LLMs, reporting wins over LoT across all nine knowledge-based VQA settings.
2026Co-authored the controlled study on spurious tool use in RL LLM agents, quantifying cue-driven shortcut behavior (up to +39.2%) and a judge-based necessity-reward fix.

Papers in the Evolve catalog

Frequent co-authors

Junda Ying×1 Zhiwei Zeng×1 Peijie Zhou×1 Zhenbin Wang×1 Lituan Wang×1 Wei Huang×1 Yan Wang×1 Zhenwei Zhang×1 Zhenbin Wang×1 Lei Zhang×1 Lituan Wang×1 Wei Huang×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.

Name resolution: 'Lei Zhang' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID or affiliation was available to confirm identity.
Institution, country, and career history are absent from the snapshot; the profile covers only the three reviewed papers and their bibliometric signals.
Output, impact, and activity dimensions are unscored because the underlying data (citation counts, venue tiers, multi-year volume) is not in the snapshot.
All reviewed papers are 2026 arXiv preprints with medium reproducibility readiness (mean 55/100); peer-review outcomes and citation records are not reflected.