Machine learning researcher focused on LLM agents — RL tool use, latent skill steering, and agentic optimization.
3
papers here
1
reviewed by Evolve
0
first-author
0
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—
Insufficient data: no publication-count or output-volume records exist in the snapshot, so this axis is unscored. It would measure the volume of papers and artifacts produced over time and its evolution.
Impact—
Insufficient data: no citation counts, h-index, or downstream-usage records are available, so impact is unscored. This axis would measure influence on later research — citations, method uptake, and reuse.
Focus100/100
Focus scores 100/100. In the bibliometric snapshot every cataloged paper maps to a single field — machine learning (natural language processing / LLM agents) — so the field distribution shows one field with no dispersion across adjacent areas. All three 2026 arXiv papers, from RL tool-use analysis to latent-skill steering and generative engine optimization, sit inside that single scope, indicating a tightly concentrated research agenda.
Collaboration60/100
Collaboration scores 60/100, based on 19 distinct co-authors in the snapshot. A recurring core cluster includes Junda Wu, Sheldon Yu, Tong Yu, Xunyi Jiang, and Yuxin Xiong, with further repeat collaborators such as David Arbour, Ritwik Sinha, and Sungchul Kim. Authorship positions are varied rather than dominant: middle author on two cataloged papers and last author on one, with no corresponding-author role recorded.
Activity—
Insufficient data: the snapshot lacks year-by-year publication records, so activity is unscored. This axis would measure recent output cadence — papers per year and continuity — not inferable from a single catalog year (2026).
Reproducibility55/100
Reproducibility scores 55/100, but this rests on a single assessed paper (reproPapers = 1), so the figure is thin. That paper — the spurious tool-use study — is rated medium readiness. Evolve's review notes the evidence is single-seed and small-scale: the 39.2% cue-driven effect and the judge-based necessity-reward fix would need multi-seed, larger-scale runs before the result can be treated as robust.
Research summary
Julian McAuley (ORCID 0000-0003-0955-7588) is profiled here from public bibliographic data and the papers Evolve has reviewed to date; the source data supplies no institution or country. The current catalog contains three 2026 arXiv papers, all situated in machine learning with a natural language processing and LLM-agents orientation — a concentration that yields a focus score of 100/100 in the underlying bibliometric snapshot. The reviewed work spans three threads: a controlled study of spurious tool use in reinforcement-learning agents, where cue-driven shortcuts inflated tool calls by 39.2% and a judge-based necessity reward corrected the behavior (rated medium reproducibility readiness, with single-seed, small-scale evidence); an empirical study of steering and optimizing latent language-model skills; and Agent2UCB, an agentic system for generative engine optimization on which he is the last-listed author. Across the snapshot he has 19 distinct co-authors, with a recurring cluster including Junda Wu, Sheldon Yu, Tong Yu, Xunyi Jiang, and Yuxin Xiong. Authorship positions are modest: middle author on two cataloged papers and last author on one, with none marked corresponding. Collaboration scores 60/100 and reproducibility 55/100, the latter resting on a single assessed paper. The output, impact, and activity axes lack the underlying data and are left unscored rather than estimated.
Collaboration pattern
The snapshot records 19 distinct co-authors. The most consistent cluster centers on Junda Wu, alongside Sheldon Yu, Tong Yu, Xunyi Jiang, and Yuxin Xiong; David Arbour, Ritwik Sinha, Sungchul Kim, Hongyi Wen, Rui Wang, Doga Dogan, and Yiwei Yang also recur. Position patterns are modest: middle author on two cataloged papers, last author on Agent2UCB, and no corresponding-author role in the reviewed set.
Research keywords
Highlights
Controlled evidence of spurious tool use in RL agents
Empirical study of latent language-model skill steering
A 2026 arXiv empirical study on steering and optimizing latent skills in language models, with McAuley as a middle author. Evolve has not yet published a verdict, keywords, or reproducibility assessment for this paper.
Agent2UCB: agentic system for generative engine optimization
A 2026 arXiv paper on an agentic system for generative engine optimization, with McAuley as last-listed author — the only cataloged paper where he holds the final author position. No Evolve verdict or reproducibility rating yet.
Timeline
- 2026Published (middle author) the controlled study 'Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act' on arXiv; Evolve's review records a 39.2% cue-driven effect corrected by a judge-based necessity reward.
- 2026Co-authored (middle author) 'Toward Latent Language Model Skills Steering and Optimization: An Empirical Study' on arXiv; not yet reviewed by Evolve.
- 2026Appeared as last author on 'Agent2UCB: Agentic System for Generative Engine Optimization' on arXiv; not yet reviewed by Evolve.
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 scores are computed only from the bibliometric snapshot supplied to this profile; institution and country are not part of the source data.
- Publication counts, citation metrics, and h-index are absent, so the output and impact axes are left unscored rather than estimated.
- The reproducibility score (55) rests on a single assessed paper (reproPapers = 1), and that study's evidence is single-seed and small-scale per Evolve's review.
- Scope is limited to three 2026 arXiv papers reviewed or indexed by Evolve; the profile is name-resolved from public bibliographic data via ORCID 0000-0003-0955-7588 and may not cover earlier or unreviewed work.