Bo Peng

Chinese Academy of Sciences · China

Unclaimed · AI-structured profileOpenAlex

CAS researcher on reward modeling and quality assessment for generative video and image models.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-09) 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 unavailable in this snapshot; the axis would measure total publication volume and the year-by-year trajectory. Only a thin base exists — 5 works in the recent three years (1 in 2024, 4 in 2026) — too little to score reliably.

Impact—

Impact scoring is not yet available; the axis would measure citation-based influence such as per-paper citations and field-weighted reach. The recorded h-index is 1 and i10-index 0, so influence cannot yet be meaningfully scored for this early-stage profile.

Focus0/100

The recorded focus score is 0, reflecting an even spread rather than one dominant cluster: fieldDist assigns 0.5 to computer vision (image quality assessment) and 0.5 to machine learning / computer vision (video generation evaluation). OpenAlex topics corroborate the breadth — Image and Video Quality Assessment (2) and Multimodal Machine Learning Applications (2), plus Advanced Neural Network Applications, Recommender Systems, and Information Retrieval (1 each). The two main threads are adjacent, both about evaluating generative visual output.

Collaboration51/100

Collaboration scores 51, a moderate level. The snapshot records 12 distinct coauthors in the last three years, and both Evolve-reviewed 2026 papers place Bo Peng in middle authorship without a corresponding role. Recurring names include Songlin Yang, Jing Dong, and Zhenchen Tang, alongside Zichuan Wang, Leilei Cao, Fengjie Zhu, Yang Li, and others — a pattern consistent with team-based CV/ML group work rather than solo output.

Activity52/100

Activity scores 52. countsByYear shows a sharp recent ramp: 1 work in 2024 rising to 4 in 2026, with 5 works across the recent three-year window. Output concentrates in 2026, when both Evolve-reviewed arXiv preprints appeared, suggesting an early but accelerating research cadence rather than a long-established record.

Reproducibility38/100

Reproducibility scores 38 (reproMean over 2 reviewed papers). RewardVerse is rated low: headline tables are missing and residual scalar instability remains despite convincing ablations. The agentic IQA paper is rated medium: its central evaluation uses a 180-image self-built set without statistical tests. Neither reviewed work yet offers a fully re-runnable evidence base.

Research summary

Bo Peng is a researcher at the Chinese Academy of Sciences working at the intersection of computer vision and machine learning, with a portfolio centered on evaluating generative visual models. The OpenAlex-resolved snapshot is early-stage: h-index 1, i10-index 0, and 5 works in the recent three-year window, with countsByYear showing 1 work in 2024 and 4 in 2026 — a steep recent ramp. Two 2026 arXiv preprints have been reviewed by Evolve, both with middle authorship. "RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling" trains video reward models with rubric-anchored optimization and GRPO to counter scalar drift; Evolve finds the ablations convincing but flags missing headline tables and residual instability (reproducibility: low). "An Evolutionary Agentic Approach for Open-ended Image Quality Perception" uses multi-agent protocols to cut MLLM holistic bias in IQA, reporting HOR falling from 44.4% to 8.6%, though the central evaluation rests on a 180-image self-built set without statistical tests (reproducibility: medium). OpenAlex topics mirror this dual thread: Image and Video Quality Assessment (2) and Multimodal Machine Learning Applications (2), plus Advanced Neural Network Applications, Recommender Systems, and Information Retrieval (1 each). Collaboration is moderate, with 12 distinct recent coauthors including Songlin Yang, Jing Dong, and Zhenchen Tang.

Collaboration pattern

Collaboration is moderate (score 51): 12 distinct coauthors in the recent three-year window. Both Evolve-reviewed 2026 papers list Bo Peng as a middle author, not corresponding. Recurring coauthors include Songlin Yang, Jing Dong, and Zhenchen Tang (who appears twice in the frequency list), plus Zichuan Wang, Leilei Cao, Fengjie Zhu, Yang Li, Xiaotong Zhao, Shuai Li, Haotian Fan, and Alan Zhao — a team-oriented pattern typical of CAS computer vision groups.

Research keywords

image quality assessmentvideo reward modelingrubric-based evaluationmulti-agent systemsmultimodal LLMreinforcement learning for video generationGRPOscalar driftholistic biastraining-free evaluationvideo generation evaluationmultimodal machine learning

Highlights

Rubric-anchored policy optimization for video reward modeling
RewardVerse trains video reward models with rubric-guided optimization and GRPO, claiming SOTA and mitigation of scalar drift via a soft-logits readout. Evolve finds the ablations convincing but notes missing headline tables and residual instability; reproducibility readiness is rated low.
Multi-agent protocols cut holistic bias in image quality assessment
This 2026 work applies an evolutionary agentic pipeline to open-ended IQA, reducing MLLM holistic bias with HOR falling from 44.4% to 8.6% in a training-free setup. Evolve notes the central evaluation relies on a 180-image self-built set without statistical tests; reproducibility readiness is medium.
A consistent thread: evaluating generative visual output
Both reviewed 2026 preprints target how to judge what generative models produce — one scoring generated video via reward modeling, the other assessing image quality with agent-built protocols. This matches the OpenAlex topic split (Image and Video Quality Assessment: 2; Multimodal Machine Learning Applications: 2) and the even 0.5/0.5 field distribution.

Timeline

2024Earliest recorded output in the OpenAlex snapshot: 1 work.
2026Output accelerates to 4 works in a single year, including two arXiv preprints reviewed by Evolve.
2026RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling posted to arXiv (middle authorship; reproducibility readiness low).
2026An Evolutionary Agentic Approach for Open-ended Image Quality Perception posted to arXiv (middle authorship; reproducibility readiness medium).

Papers in the Evolve catalog

Frequent co-authors

Songlin Yang×2 Jing Dong×2 Zhenchen Tang×1 Zichuan Wang×1 Leilei Cao×1 Fengjie Zhu×1 Zhenchen Tang×1 Yang Li×1 Xiaotong Zhao×1 Shuai Li×1 Haotian Fan×1 Alan 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.

Bibliometric metrics come from an OpenAlex-resolved identity; no ORCID is linked, so name ambiguity is possible.
Output and impact dimensions are unscored in this snapshot; h-index 1 and i10-index 0 are too early to interpret as influence.
Qualitative verdicts draw on only 2 Evolve-reviewed papers, both 2026 arXiv preprints; peer-reviewed venues are not yet reflected.
The reproducibility score (38) covers just 2 papers and should not be read as a general quality verdict.