LLM agent safety researcher at Fudan University: benchmarking agent conflicts, co-evolving harness-policy safety alignment
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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: this profile lacks a reliable count of total publications and venues, since the visible record is limited to a small set of 2026 arXiv works. This axis would measure cumulative publication volume across the career.
Impact—
Not yet scored: citation influence cannot be computed — h-index and i10-index are recorded as 0 and no citation aggregates are available, consistent with works that all appeared in 2026. This axis would measure citation uptake and scholarly influence.
Focus0/100
The field distribution splits evenly between machine learning–oriented LLM agent safety alignment (0.5) and AI safety / LLM agent security (0.5), so the two reviewed papers sit within one broad problem area. Topic counts, however, spread across five labels — adversarial robustness (3), ethics and social impacts of AI (2), safety systems engineering in autonomy (2), reinforcement learning in robotics (1), and robot manipulation and learning (1) — and the computed focus score is 0 in this snapshot.
Collaboration58/100
The collaboration score is 58, built on 17 distinct coauthors in the available record. Repeated collaborators include Xia Hu, Qinghua Mao, Wanying Qu, Dadi Guo, Dongrui Liu, and Yuejin Xie. Both reviewed papers list this researcher in a middle author position rather than as corresponding author, consistent with contribution within a larger LLM agent safety group at Fudan University.
Activity47/100
Activity scores 47. The counts-by-year record shows 4 works in 2026, matching the 4 works in the most recent three-year window — an output rhythm concentrated entirely in the current year, with no earlier years present in this snapshot. This suggests an early or newly visible phase of publishing rather than an established multi-year cadence.
Reproducibility78/100
Reproducibility is the strongest recorded axis at 78, computed from 2 reviewed papers. ClashBench carries high reproducibility readiness as an executable benchmark, while SafeEvolve is rated medium: its harness-policy co-evolution results are broad and well-ablated, but single-run numbers lack significance testing. Together they indicate strong artifact and protocol disclosure for this career stage.
Research summary
Yanwei Fu is a researcher at Fudan University (China) whose visible record centers on the safety of large language model (LLM) agents. The available field distribution splits evenly between machine learning–oriented LLM agent safety alignment (0.5) and AI safety / LLM agent security (0.5), and OpenAlex records 4 works in 2026 — the entire output in the current snapshot. Two of these have been reviewed by Evolve. The first, ClashBench (2026), is an executable benchmark reporting that LLM agents destructively preempt incumbent tasks in 44.5% of runs, with the caveat that single-run statistics merit caution. The second, SafeEvolve (2026), proposes harness-policy co-evolution from agent experience for safety alignment, using reinforcement learning (GRPO) and addressing indirect prompt injection; it claims roughly 3x reduction in agent attack success, with broad, well-ablated results but single-run numbers lacking significance testing. Topic counts concentrate on adversarial robustness in machine learning (3), ethics and social impacts of AI (2), and safety systems engineering in autonomy (2), with single mentions in reinforcement learning in robotics and robot manipulation and learning. Reproducibility is the strongest recorded dimension (78 across 2 papers), while collaboration scores 58 over 17 distinct coauthors and activity 47. Citation metrics (h-index, i10-index) are not yet populated, so output and impact remain unscored in this profile.
Collaboration pattern
With 17 distinct coauthors and a collaboration score of 58, this researcher works embedded in a sizeable Fudan University LLM agent safety group. Recurring collaborators include Xia Hu, Qinghua Mao, Wanying Qu, Dadi Guo, Dongrui Liu, and Yuejin Xie. Both catalog papers place this researcher in a middle author position without corresponding-author status, suggesting a hands-on technical contributor within team-led projects rather than an independent group lead so far.
Research keywords
Highlights
ClashBench quantifies destructive preemption by LLM agents
SafeEvolve co-evolves harness and policy for safety alignment
A measure-then-mitigate agenda for LLM agent safety
The two reviewed 2026 papers pair an executable conflict benchmark with an alignment method, jointly targeting agentic side effects, resource conflicts, and indirect prompt injection — with reproducibility readiness rated high and medium respectively.
Timeline
- 2026Co-authored ClashBench, an executable benchmark reporting that LLM agents destructively preempt incumbent tasks in 44.5% of runs.
- 2026Co-authored SafeEvolve, proposing harness-policy co-evolution for safety alignment with a claimed roughly 3x reduction in agent attack success.
- 2026OpenAlex counts-by-year records 4 works in 2026, the only publication year visible in this snapshot.
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.
- Bibliometric signals are computed from OpenAlex only (identity: openalex); no ORCID was available for disambiguation.
- 'Yanwei Fu' is a common Chinese name; this profile is name-resolved from public bibliographic data and may not capture all works by same-named authors.
- Citation-based metrics (h-index 0, i10-index 0) are unpopulated; the output and impact dimensions are therefore not scored.
- Only 2 Evolve-reviewed papers, both 2026 arXiv preprints, underpin the reviewed findings; other works in the record have not yet been assessed.