AI safety researcher at Shanghai AI Laboratory: executable LLM-agent safety benchmarks and harness-policy alignment
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Citation-record output is not yet available: the OpenAlex snapshot lists no aggregated publication totals for this author, so research volume cannot be scored. This axis would measure the number and growth trajectory of the author's outputs over time.
Impact—
Impact data is insufficient in this snapshot: no aggregated citation counts or h-index totals are available, so influence cannot be scored here. This axis would measure citations, uptake, and the scholarly reach of the author's published work.
Focus0/100
Focus scores 0 because the record splits evenly (50/50) between two adjacent fields — LLM agent safety alignment and AI safety / LLM agent security — across the 2026 works. Topically, however, the six OpenAlex topic assignments cluster tightly: adversarial robustness (2) alongside multi-agent negotiation, AI ethics, access control, and verification (1 each) read as one coherent agent-safety agenda.
Collaboration61/100
Collaboration scores 61, built on 20 distinct coauthors across a three-paper record — a moderately broad network for such an early snapshot. Twelve recurring partners appear, including Yanwei Fu, Lin Chen, and Ziyi Liu. On all cataloged papers Xia Hu holds a middle author position rather than first or corresponding, consistent with contributing within larger team efforts.
Activity40/100
Activity scores 40: recent3yWorks is 3 and countsByYear shows all three papers dated 2026, with no earlier years recorded. Total volume is small, but the output is entirely current — a newly active record whose trajectory three data points cannot yet establish.
Reproducibility78/100
Reproducibility scores 78, the mean across the two Evolve-rated papers. ClashBench rates high — an executable benchmark with concrete run-level results (44.5% destructive preemption) — while SafeEvolve rates medium: broad and well-ablated but single-run numbers without significance testing. A strong artifact orientation for such an early record.
Research summary
Xia Hu is an AI safety researcher at Shanghai Artificial Intelligence Laboratory whose cataloged work centers on the security and alignment of LLM-based agents. The Evolve catalog holds three 2026 arXiv papers, each with Xia Hu in a middle author position. ClashBench introduces an executable benchmark showing that LLM agents destructively preempt incumbent tasks in 44.5% of runs, quantifying agentic side effects under resource conflicts. SafeEvolve proposes harness-policy co-evolution from agent experience, applying reinforcement learning (GRPO) to defend against indirect prompt injection, with a reported ~3x reduction in agent attack success. A third paper examines how AI agents reshape consensus formation in human groups, extending the agenda toward AI's social impact. Bibliometrically the record is compact and recent: three works, all dated 2026, split evenly (50/50) between two adjacent fields — machine learning for LLM agent safety alignment and AI safety / LLM agent security — with adversarial robustness the most frequent OpenAlex topic (2 of 6 topic assignments). Xia Hu collaborates with 20 distinct coauthors, including recurring partners such as Yanwei Fu, Lin Chen, and Ziyi Liu. Reproducibility signals are comparatively strong, averaging 78 across two rated papers. Citation-based output and impact metrics are not yet available in this snapshot.
Collaboration pattern
Xia Hu's OpenAlex record lists 20 distinct coauthors across three 2026 papers, with 12 recurring partners — Yanwei Fu, Lin Chen, Ziyi Liu, Yong Li, Qinghua Mao, Wanying Qu, Dadi Guo, Leitao Yuan, Qingyu Liu, Yu Li, Guanxu Chen, and Xi Lin. On every cataloged paper Xia Hu is a middle author rather than first or corresponding, indicating contributions embedded in team-led projects rather than solo or PI-driven output.
Research keywords
Highlights
ClashBench: executable benchmark of destructive agent preemption
SafeEvolve: harness-policy co-evolution for agent safety alignment
AI agents reshape consensus formation in human groups
A third 2026 arXiv paper, cataloged without field tags, keywords, an Evolve review, or a reproducibility rating yet, extends the record from agent security toward AI's social and behavioral influence on group decision-making.
Timeline
- 2026Publishes ClashBench, an executable benchmark in which LLM agents destructively preempt incumbent tasks in 44.5% of runs.
- 2026Publishes SafeEvolve, proposing harness-policy co-evolution trained with reinforcement learning (GRPO) for agent safety alignment.
- 2026Co-authors a study on how AI agents reshape consensus formation in human groups.
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 come from OpenAlex only; no ORCID is linked, and 'Xia Hu' is a common Chinese name, so this profile is name-resolved from public bibliographic data.
- The record contains only three 2026 arXiv preprints; citation counts, venue acceptance, and peer-reviewed publication history are absent from the snapshot.
- h-index and i10-index are recorded as 0, so the output and impact radar axes are left unscored rather than estimated.
- The third paper (AI agents reshape consensus formation in human groups) has no Evolve review, keywords, or reproducibility rating yet.