Yangxu Liao

Wuhan University · China

Unclaimed · AI-structured profileOpenAlex

LLM agent evaluation and recursive self-improvement researcher at Wuhan University; profile name-resolved from OpenAlex

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Output data is insufficient: only 5 recent works are visible and the full publication record is unresolved under this name, so no score is assigned. This axis would measure publication volume, venue quality, and output consistency over time.

Impact—

Impact data is insufficient: the record shows h-index 0 and i10-index 0 with no usable citation counts, so no score is assigned. This axis would measure citation-based influence, including citation counts, h-index trajectory, and uptake of the 2026 preprints.

Focus0/100

The focus score of 0 reflects a highly dispersed topical signature. The field distribution splits evenly at 0.5 and 0.5 between software engineering (LLM agent evaluation) and machine learning (recursive self-improvement of LLM systems), while OpenAlex topics add three parallel clusters — generative adversarial networks and image synthesis, video analysis and summarization, and digital media forensic detection, at 4 works each — plus smaller clusters in AI planning and evolutionary algorithms.

Collaboration74/100

Collaboration scores 74, supported by 40 distinct coauthors in the record. The frequent-coauthor list includes Guancheng Wan, Jiajun Wu, Leixin Sun, Zihan Tan, Yitao Liu, Shuo Li, Jiaru Qian, Shanghaoran Quan, Chuangxin Zhao, Yang Liu, and Bin Chong. Both reviewed 2026 papers list this researcher in a middle author position, indicating work within larger team efforts rather than solo or corresponding-author leadership.

Activity52/100

Activity scores 52, based on a recent three-year window containing 5 works, all dated 2026 in the counts-by-year record. The visible output is therefore concentrated in a single year, with no dated works in earlier years in this snapshot. Both catalog papers are 2026 arXiv preprints, consistent with a newly visible or early-stage publication record whose cadence cannot yet be characterized across multiple years.

Reproducibility55/100

Reproducibility scores 55, the mean across 2 reviewed papers, both rated medium reproducibility readiness. SWE-Prometheus earns credit for rigorous controls — no-op, template, and gate ablations plus mutation testing — though single-rollout LLM-judged scoring caps precision; MetaRSI/RSI2 shows serious engineering but keeps all validation inside the closed-format slice it critiques. The moderate score reflects solid methodological hygiene offset by evaluation-pipeline limitations.

Research summary

Yangxu Liao is a researcher at Wuhan University whose profile, name-resolved from OpenAlex bibliographic data, currently centers on evaluation and self-improvement of large language model (LLM) agents. The field distribution splits evenly (0.5/0.5) between software engineering — LLM agent evaluation — and machine learning — recursive self-improvement of LLM systems. Two 2026 arXiv papers reviewed by Evolve illustrate this: SWE-Prometheus, a benchmark measuring engineering-governance improvements in real repositories with rigorous no-op, template, and gate-ablation controls, and MetaRSI/RSI2, a meta-level scheduler that budgets data, harness, and weight self-improvement operators. Both list the author in a middle position. OpenAlex topics add a computer-vision signature — generative adversarial networks and image synthesis, video analysis and summarization, and digital media forensic detection (4 works each) — which diverges from the LLM-agent line and warrants caution. Quantitative signals are early-stage: 5 works in the recent three-year window (all dated 2026), 40 distinct coauthors, and h-index and i10-index of 0, so output and impact are left unscored. Radar scores show collaboration at 74, activity at 52, reproducibility at 55 (mean across 2 reviewed papers, both at medium reproducibility readiness), and focus at 0, reflecting a dispersed topical footprint.

Collaboration pattern

The collaboration score of 74 rests on 40 distinct coauthors in the OpenAlex record. Recurring names include Guancheng Wan, Jiajun Wu, Leixin Sun, Zihan Tan, Yitao Liu, Shuo Li, Jiaru Qian, Shanghaoran Quan, Chuangxin Zhao, Yang Liu, and Bin Chong. Both reviewed 2026 papers are middle-author contributions to team projects spanning LLM agent evaluation and recursive self-improvement, suggesting a collaborative working style within larger groups rather than solo or corresponding-author-led output.

Research keywords

LLM coding agentsAgent benchmarkingSoftware engineering governanceRepository-level evaluationRecursive self-improvementHarness / scaffold evolutionSynthetic data generationSelf-play evaluationMeta-learning schedulersGenerative adversarial networksDigital media forensicsVideo analysis and summarization

Highlights

SWE-Prometheus: benchmarking LLM agents on real-repository engineering governance
A 2026 arXiv benchmark measures whether LLM coding agents genuinely improve governance in real-world repositories, with rigorous no-op, template, and gate-ablation controls plus mutation testing; Evolve notes that single-rollout LLM-judged scoring caps precision, and reproducibility readiness is medium.
MetaRSI/RSI2: budgeting data, harness, and weight self-improvement under one scheduler
A 2026 arXiv system schedules data, harness, and weight self-improvement operators for recursive self-improving LLM systems under a single budget; Evolve finds the engineering serious but all validation confined to the closed-format slice the work itself critiques. Reproducibility readiness: medium.

Timeline

2026SWE-Prometheus preprint released on arXiv, introducing governance-improvement benchmarks for LLM coding agents with no-op, template, and gate-ablation controls.
2026MetaRSI/RSI2 preprint released on arXiv, scheduling data, harness, and weight self-improvement operators for recursive self-improving systems under one budget.
2026OpenAlex snapshot records 5 works dated 2026 and 40 distinct coauthors, with h-index and i10-index at 0 — an early-stage, team-heavy record.

Papers in the Evolve catalog

Frequent co-authors

Guancheng Wan×2 Jiajun Wu×1 Leixin Sun×1 Zihan Tan×1 Yitao Liu×1 Shuo Li×1 Jiaru Qian×1 Shanghaoran Quan×1 Chuangxin Zhao×1 Yang Liu×1 Bin Chong×1 Zihan Tan×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.

Bibliometrics come from OpenAlex only; no ORCID is provided, so this profile is name-resolved and may merge or split namesakes.
The OpenAlex topic profile (generative models, video summarization, media forensics) diverges from the LLM-agent fields of the reviewed papers; field attribution should be read with caution.
h-index and i10-index are 0 and no citation counts are available, so output and impact are left unscored rather than estimated.
The reproducibility score covers only the 2 papers reviewed by Evolve; other outputs were not assessed.