Early-career Chinese Academy of Sciences researcher spanning RL post-training, diffusion LMs, and agentic AI evaluation
4
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
4
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: the OpenAlex snapshot carries no aggregate publication-volume metric beyond per-year counts, so the output axis — which would measure total scholarly production and its growth trajectory — is left unscored rather than estimated.
Impact—
Not yet scored: citation-side signals are too thin in this snapshot (h-index 4, i10-index 0), so impact — which would measure community uptake, citation accumulation and influence of the work — is not estimated.
Focus0/100
The focus score is 0 because output splits perfectly evenly across four subfields — GUI agents/vision-language models, multimodal AI agents, LLM agent evaluation, and diffusion-LM post-training each take 25% of the field distribution — while OpenAlex topics range from multimodal ML applications (28 works) to topic modeling (15) and smaller clusters in handwritten text recognition, image/video retrieval and GAN synthesis (5 each). The pattern reads as breadth-first, theme-hopping early output rather than a settled specialization.
Collaboration73/100
With 37 distinct coauthors across 37 works in the trailing three-year window, the collaboration score is 73. Recurring names include Zhenbo Luo (listed twice), Shaojie Zhang, Jian Luan, Yan Zhang, Daiqing Wu and Yu Zhou. On the four Evolve-reviewed papers Pei Fu sits in middle or last author positions and is corresponding author on none, consistent with a rotating-team contributor pattern rather than a fixed project-lead role so far.
Activity100/100
Activity scores 100: OpenAlex counts by year rise from 3 works (2024) to 9 (2025) and 25 (2026), with 37 works in the trailing three-year window. All four Evolve-reviewed papers are 2026 arXiv preprints, confirming an intense, currently peaking publication cadence typical of a fast-moving early-career phase.
Reproducibility38/100
Reproducibility scores 38, the mean readiness across the 4 papers Evolve has assessed. Reviews repeatedly flag missing uncertainty reporting — single-seed reinforcement learning runs without confidence intervals, absent significance tests, single-run benchmark results, and a thin 2.03-point margin over one reproduced baseline — with two papers rated only medium readiness and two rated low.
Research summary
Pei Fu is an early-career researcher at the Chinese Academy of Sciences whose recent work concentrates on training and evaluating large-model agents. The bibliometric record is young but steep: 3 works in 2024, 9 in 2025 and 25 in 2026, with 37 works and 37 distinct coauthors in the trailing three-year window and an h-index of 4. Four 2026 arXiv preprints reviewed by Evolve sketch the agenda: informed masking for reinforcement learning in diffusion language models (built on LLaDA-8B across three RL frameworks), masked trajectory prediction for omni-dimensional GUI agent navigation, NTEP-R process rewards for tool use in agentic vision-language models, and DAREBench, a deployment-aware benchmark unifying 233 tasks and 35 models in a shared OpenClaw runtime with evidence-audited scoring. Across these, Evolve's verdicts consistently credit promising, consistent gains while flagging thin statistical reporting — single-seed runs, no confidence intervals, no significance tests — yielding a mean reproducibility readiness of 38. Field distribution is perfectly even across four subfields (GUI agents, multimodal agents, agent evaluation, and diffusion-LM post-training at 25% each), signaling an exploratory phase rather than settled specialization. Collaboration roles so far are middle and last authorship without corresponding-author positions, and publication output is at its recorded peak.
Collaboration pattern
Pei Fu publishes in rotating team settings: 37 distinct coauthors over 37 recent works and a collaboration score of 73. Zhenbo Luo appears twice in the frequent-coauthor list; other recurring names include Shaojie Zhang, Jian Luan, Yan Zhang, Daiqing Wu, Huawen Shen, Ruoceng Zhang, Jiahui Yang, Yu Zhou and Can Ma. On the four reviewed papers Pei Fu holds middle or last author positions, none as corresponding author — a contributor rather than project-lead pattern to date.
Research keywords
Highlights
Informed Masking: structure-aware perturbation for diffusion-LM RL
Masked trajectory prediction unifies GUI agent navigation
DAREBench: deployment-aware agent evaluation at scale
NTEP-R: tool-evidence path rewards for agentic vision-language models
Timeline
- 2024First recorded output in the OpenAlex record: 3 works appear in 2024, an early-career starting point.
- 2025Output grows to 9 works as the multimodal-agent research direction takes shape across collaborating teams.
- 2026Peak year: 25 works recorded, including four Evolve-reviewed arXiv preprints covering diffusion-LM RL, GUI agent navigation, agent evaluation (DAREBench) and tool-use process rewards for vision-language agents.
Papers in the Evolve catalog
Informed Masking: Structure-Aware Perturbation for Reinforcement Learning in Diffusion Large Language Models
Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction
DAREBench: Deployment-Aware and Reliable Evaluation of Models as Agents
Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models
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.
- Bibliometrics come from OpenAlex only; no ORCID is linked and Pei Fu is a common Chinese name — the frequent-coauthor list even contains a separate Pei Fu entry, so identity resolution may carry disambiguation noise.
- Output and impact axes are unscored because citation-side signals are too thin in the snapshot (h-index 4, i10-index 0).
- All four reviewed papers are 2026 arXiv preprints; peer-reviewed venue status is not captured in the data.
- The reproducibility score averages Evolve reviews of only 4 papers (mean 38), a small sample.