Yuhao Wu

Yangtze River Delta Physics Research Center (China) · Singapore

Unclaimed · AI-structured profileOpenAlex

LLM safety and self-evolution researcher building benchmarks and defenses for self-improving language models

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: OpenAlex lists 13 recent works but volume- and venue-normalized productivity cannot be computed from this snapshot. This axis would measure annual publication volume and venue quality.

Impact—

Impact scoring is unavailable: the snapshot reports h-index 0 and i10-index 0 with no citation counts by year, so citation influence cannot yet be assessed. This axis would measure citations, citation velocity, and downstream adoption of the work.

Focus100/100

The focus score of 100 reflects an exceptionally concentrated recent record: of 13 works in the last three years, topic assignments cluster on Topic Modeling (5) with Natural Language Processing Techniques, Software Engineering Research, Fault Detection and Control Systems, and AI-based Problem Solving and Planning at 2 each, while the reviewed output funnels into a single flagged field, AI safety / LLM security, across LLM defense and self-evolution benchmarks.

Collaboration71/100

Collaboration scores 71 on the strength of 35 distinct coauthors across 13 recent works and a recurring core of 12 frequent collaborators, including Jiajun Shi, Yuxuan Zhang, and Ge Zhang. The reviewed papers show team-based arXiv output with one first-author role (Aspire) and middle positions elsewhere, consistent with an early-career contributor embedded in larger LLM research groups.

Activity77/100

Activity scores 77, driven by a steep recent ramp: 13 works in the last three years, of which 12 appeared in 2026 versus only 1 in 2025 — a more than tenfold year-over-year jump. Three of these works were reviewed by Evolve in 2026, all arXiv preprints spanning LLM security and agent self-improvement benchmarks, indicating an intensive ongoing production cycle.

Reproducibility55/100

Reproducibility scores 55 (reproMean 55 over 1 scored paper). Among the reviewed works, Bait-and-Recover and S3Gym carry medium reproducibility readiness while Aspire is rated low. Evolve's verdicts flag methodological caution — only the default estimator tested in Bait-and-Recover and 3-episode evaluations in S3Gym — suggesting artifacts exist but statistical rigor limits independent replication.

Research summary

Yuhao Wu (Yangtze River Delta Physics Research Center) is an early-career researcher working at the intersection of LLM safety and LLM self-evolution. The OpenAlex snapshot attributes 13 works in the last three years — 12 in 2026 and 1 in 2025 — with h-index and i10-index both currently 0, so citation impact cannot yet be scored. Topic modeling dominates the topic profile (5 topic assignments), followed by natural language processing techniques, software engineering research, fault detection and control systems, and AI-based problem solving and planning (2 each). Within the Evolve-reviewed catalog, three 2026 arXiv papers define the core agenda: Aspire (first author) benchmarks whether models can self-evolve from vague goals under hidden expert evaluation; S3Gym (middle author) asks whether self-testing and self-judging translate into self-improvement; and Bait-and-Recover (middle author) defends LLMs against white-box editing jailbreaks by poisoning internal refusal signals, lifting minimum refusal from 16% to 72% at KL divergence at most 0.10. Both self-evolution benchmarks report honest but statistically thin negative results. The record shows high topical focus (radar 100), strong recent activity (77), broad collaboration across 35 distinct coauthors (71), and medium reproducibility (55), positioning this as an early-stage but rapidly expanding profile.

Collaboration pattern

Yuhao Wu works in sizeable teams: 35 distinct coauthors across 13 recent works, with a recurring core of 12 collaborators including Jiajun Shi, Yuxuan Zhang, Xinping Lei, Zexuan Wang, Ge Zhang, and Jiaheng Liu. Authorship in the reviewed papers spans first-author leadership (Aspire) and middle positions (S3Gym, Bait-and-Recover), with no corresponding-author role recorded — a pattern typical of an early-career researcher contributing to shared benchmarks and defenses.

Research keywords

LLM safety alignmentWhite-box jailbreak defenseRepresentation engineeringRefusal directionLLM self-evolutionSelf-improving agentsAgent benchmarkingSelf-judgingAutonomous post-trainingVague goalsTopic modelingIn-context learning

Highlights

Refusal-signal poisoning blunts white-box editing jailbreaks
Co-authored work proposing bait/recovery adapters that poison an attacker's refusal-direction readout, raising minimum refusal from 16% to 72% at KL divergence at most 0.10; Evolve notes only the default estimator was tested.
First-author benchmark asks whether models self-evolve from vague goals
As first author, built a benchmark for vague-goal LLM self-evolution with a hidden expert evaluation set; results are honest but statistically thin — self-evolution gains are rare and the evolved harness remains below the reference.
S3Gym tests whether self-judging becomes self-improvement
Co-authored benchmark showing that self-improvement from self-testing and self-judging is task-dependent; the protocol is sound and results honest, but tiny 3-episode evaluations lack statistical rigor.

Timeline

2025First recorded output in the OpenAlex snapshot: 1 work appears in 2025, before the 2026 ramp-up.
2026Output accelerates to 12 works in a single year, establishing the LLM safety and self-evolution research agenda.
2026Publishes Aspire as first author, benchmarking whether LLMs can self-evolve from vague goals under hidden expert evaluation.
2026Co-authors S3Gym, showing that self-testing and self-judging translate into self-improvement only in a task-dependent way.
2026Co-authors Bait-and-Recover, defending LLMs against white-box editing jailbreaks by poisoning internal refusal signals (minimum refusal 16% to 72%).

Papers in the Evolve catalog

2026-09-05
Bait-and-Recover: Poisoning Internal Refusal Signals to Defend LLMs against White-Box Editing Jailbreaks
2026-08-31first author
Aspire: Can Models Self-Evolve from Vague Goals?
2026-08-31
S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?

Frequent co-authors

Jiajun Shi×2 Yuxuan Zhang×2 Xinping Lei×2 Zexuan Wang×2 Jinkai Liu×2 Jiaheng Liu×2 Shen Yan×2 Wenhao Huang×2 Ge Zhang×2 Jingyuan Zhang×1 Junting Zhou×1 Yuchen Wu×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.

Name 'Yuhao Wu' is common; this profile is name-resolved from public OpenAlex bibliographic data and may merge or split identities.
Bibliometrics come from a single OpenAlex snapshot; h-index and i10-index are both 0 and citation counts are absent, so impact is not scored.
Affiliation metadata lists the Yangtze River Delta Physics Research Center (China) while the country code is SG; the discrepancy is unresolved in the source data.
The qualitative assessment rests on only three Evolve-reviewed papers (all 2026 arXiv preprints); peer-reviewed venues and the full publication history are not covered.