Zhaolu Kang

Peking University · China

Unclaimed · AI-structured profileOpenAlex

Peking University researcher in multimodal in-context learning and reward-aligned post-training for LLM reasoning

papers here

reviewed by Evolve

first-author

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 scoring is not yet available; this axis would measure publication volume and productivity trends over time. The record shows an h-index of 3 and i10-index of 0 alongside 59 recent works — too early for a stable score.

Impact—

Impact scoring is unavailable; it would capture citation influence such as citation counts and h-index trajectory. With an i10-index of 0 and h-index of 3, the citation base is too thin to score reliably at this stage.

Focus0/100

Focus scores 0 because the OpenAlex field distribution splits exactly evenly — 50% multimodal machine learning (vision-language models) and 50% machine learning LLM post-training/reinforcement learning — rather than concentrating in one field. Topic assignments reinforce the spread: topic modeling (30 works) and multimodal ML applications (29) lead, followed by domain adaptation and few-shot learning (13), NLP techniques (12), and reinforcement learning in robotics (6), indicating breadth across at least two active research lines.

Collaboration73/100

Collaboration scores 73, supported by 37 distinct coauthors over the last three years and 12 recurring collaborators, including Leqi Zheng, Richeng Xuan, Zhichao Hu, and Eric Hanchen Jiang. Both Evolve-reviewed 2026 papers are mid-author team efforts, while the two first-author preprints suggest the researcher also leads writing within a large, active collaboration network centered on Peking University.

Activity100/100

Activity reaches the maximum score of 100: the counts-by-year record rises from 2 works in 2024 to 5 in 2025 and 52 in 2026, totaling 59 works in the last three years. All four catalog papers appeared as 2026 arXiv preprints, confirming an intense and ongoing submission cadence.

Reproducibility55/100

Reproducibility scores 55, computed from 2 Evolve-reviewed papers with a mean readiness of 55, both rated medium. The COMIL review notes strong results — winning 13/16 MLLM-dataset settings — but flags the absence of significance tests and released code; the gradient-aligned-rewards paper reports gains at under 9% overhead that have yet to be independently replicated.

Research summary

Zhaolu Kang is a researcher at Peking University whose profile is name-resolved from OpenAlex bibliographic data. The record is early but fast-moving: an h-index of 3 and i10-index of 0 sit alongside 59 works in the last three years, including 52 indexed in 2026 alone — a steep ramp from 2 works in 2024 and 5 in 2025. Topic mass concentrates in topic modeling (30 works) and multimodal machine learning applications (29), with secondary clusters in domain adaptation and few-shot learning (13), NLP techniques (12), and reinforcement learning in robotics (6); the field distribution splits exactly evenly between vision-language multimodal ML and LLM post-training with reinforcement learning. Four 2026 arXiv preprints anchor the Evolve catalog. Two first-author papers — JPO for structured legal reasoning in criminal judgment prediction, and Modality Fault Lines on fragile omni-modal reasoning under structural corruptions — mark an emerging independent agenda. Two mid-author papers carry full Evolve reviews: COMIL, which contrasts suboptimal and better demonstrations for multimodal in-context learning and wins 13/16 MLLM-dataset settings, and gradient-aligned rewards that lift RLVR math pass@k at under 9% overhead. Work with 37 distinct coauthors indicates strongly team-based research; reproducibility evidence remains partial (mean 55 across 2 reviewed papers).

Collaboration pattern

The record lists 37 distinct coauthors over the last three years and 12 recurring collaborators — Leqi Zheng, Richeng Xuan, Zhichao Hu, Yantao Liu, Tailong Luo, Lei Wei, Chenghua Zhu, Junhao Gong, Jiachen Qian, Eric Hanchen Jiang, Jiaxin Liu, and Yuan Wang. The two Evolve-reviewed 2026 papers are mid-author team efforts, while JPO and Modality Fault Lines are first-author, suggesting leadership within a large, Peking University-centered network rather than solo work.

Research keywords

Multimodal in-context learningMultimodal large language modelsContrastive demonstration selectionDemonstration retrievalReasoning path alignmentReinforcement learning with verifiable rewards (RLVR)Reward shapingGradient alignmentProcess reward modelingOmni-modal robustnessLegal judgment predictionDomain adaptation and few-shot learning

Highlights

COMIL: contrastive demonstration modeling for multimodal in-context learning
Per Evolve's review, COMIL contrasts suboptimal versus better demonstration responses to align reasoning paths in multimodal in-context learning, winning 13 of 16 MLLM-dataset settings; the review notes strong results but no significance tests or released code.
Gradient-aligned rewards for RLVR math reasoning
This mid-author 2026 preprint proposes expert chain-of-thought gradient-cosine rewards for RLVR training, improving math pass@k at under 9% overhead; Evolve finds the theory well-motivated but the gains small on the tiny benchmarks used.
JPO: first-author policy optimization for legal judgment prediction
First-author 2026 arXiv preprint proposing Juris Policy Optimization for structured legal reasoning in criminal judgment prediction; it is not yet reviewed by Evolve, so no verdict or reproducibility rating is available.
Modality Fault Lines: stress-testing omni-modal reasoning
First-author 2026 arXiv preprint showing that structural corruptions expose fragile reasoning in omni-modal models; it awaits Evolve review, with no keywords or reproducibility rating yet.

Timeline

2024Earliest indexed output year in the OpenAlex snapshot, with 2 works recorded.
2025Output grows to 5 indexed works, an early but modest expansion of the record.
2026Output surges to 52 indexed works (59 across 2024-2026); four arXiv preprints enter the Evolve catalog, including first-author JPO on legal judgment prediction and Modality Fault Lines on omni-modal robustness.

Papers in the Evolve catalog

Frequent co-authors

Leqi Zheng×2 Richeng Xuan×2 Zhichao Hu×2 Yantao Liu×1 Tailong Luo×1 Lei Wei×1 Chenghua Zhu×1 Junhao Gong×1 Jiachen Qian×1 Eric Hanchen Jiang×1 Jiaxin Liu×1 Yuan Wang×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 an OpenAlex-linked record (identity: openalex); no ORCID is attached, so this profile is name-resolved from public bibliographic data and may merge or split same-name authors.
Output and impact dimensions lack underlying scoring data; the reported h-index (3) and i10-index (0) indicate a very early citation record.
Only 2 of the 4 catalog papers carry full Evolve reviews, so the reproducibility mean (55) rests on those two papers alone.
All four catalog papers are 2026 arXiv preprints; peer-reviewed venue status is not captured in the provided data.