Yu Liu

—

Unclaimed · AI-structured profileOpenAlex

LLM agent evaluation across games, education, and multimodal agents — testing whether models reason or just extrapolate

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-09-20) 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: no publication-volume or productivity data is available for this profile. This axis would measure total output over time — paper counts by venue and year — once a fuller bibliographic record is linked.

Impact—

Not yet scored: citation counts, h-index, and downstream influence are unavailable in the current data. This axis would measure how widely the three 2026 preprints are cited and taken up by subsequent work.

Focus0/100

Focus scores 0 because the field distribution splits exactly into thirds: multi-agent LLM evaluation with behavioral game theory, NLP/AI in education, and multimodal agent evaluation each account for 0.333 of the record. Rather than a narrow specialty, the profile shows a deliberately spread evaluation agenda — three domains, three first-author papers — trading depth of concentration for breadth across LLM evaluation settings.

Collaboration60/100

Collaboration scores 60, built on 19 distinct coauthors across the three first-author 2026 papers. Recurring names include Wenwen Li, Yifan Dou, Guangnan Ye, Zeming Liu, Tianle Zhang, Zihao Cheng, Yuhang Guo, Kehai Chen, Min Zhang, Yunhong Wang, and Haifeng Wang, suggesting stable small teams rather than one-off pairings; all three papers keep Liu in the first author position.

Activity—

Not yet scored: the data covers only the three catalogued 2026 preprints, so recent publication cadence cannot be computed. This axis would measure publishing tempo and ongoing research activity across years.

Reproducibility55/100

Reproducibility scores 55, the mean across three assessed papers (reproMean 55, reproPapers 3), each rated medium readiness. All three are arXiv preprints with concrete artifacts — PersonaPath's 2,000-persona, 347-textbook benchmark and MIRAGE's six artifacts — though Evolve's reviews note simulation-only evaluation and a single runtime as limits on generality.

Research summary

Yu Liu is an early-career researcher whose 2026 arXiv output approaches one question from three directions: when do large language model agents actually reason, and how should we evaluate them? All three catalogued papers are first-author preprints. In multi-agent behavioral game theory, Liu asks whether in-context learning reflects recursive belief reasoning or statistical extrapolation, using a rational-expectations-equilibrium diagnostic in public goods and global game settings; gains vanish when price trends break, though the reasoning null is not airtight. In AI in education, PersonaPath introduces a knowledge-centric benchmark for personalized learning path planning built from 2,000 personas and 347 textbooks, where the best model passes only 29.5% and evaluation remains simulation-only. In multimodal agent evaluation, MIRAGE shows that pre- and post-compaction conversation states form distinct grounding-failure regimes for historical evidence use, though six artifacts and one runtime bound generality. The evenly split three-field distribution yields a focus score of 0, reproducibility readiness averages 55 across three medium-rated papers, and collaboration scores 60 with 19 distinct coauthors. Output, impact, and activity metrics are not yet available. Together, the record reads as a methods-first evaluation agenda rather than a single domain application.

Collaboration pattern

Yu Liu's three first-author 2026 papers draw on 19 distinct coauthors (collaboration score 60). Recurring collaborators include Wenwen Li, Yifan Dou, Guangnan Ye, Zeming Liu, Tianle Zhang, Zihao Cheng, Yuhang Guo, Kehai Chen, Min Zhang, Yunhong Wang, and Haifeng Wang; the frequent-coauthor list also contains a 'Yu Liu' entry, most plausibly a same-name artifact. The pattern suggests small rotating teams across three evaluation domains rather than one long-running group.

Research keywords

in-context learningLLM agent evaluationbehavioral game theoryrational expectations equilibriumglobal gamesLLM benchmarkingpersonalized educationlearning path planningknowledge graphsmultimodal agentscontext compactionlong-context evaluation

Highlights

Reasoning or extrapolation? A diagnostic for LLM agents in games
First-author study asking whether in-context learning in multi-agent interdependent decision-making reflects recursive belief reasoning or statistical extrapolation. Using a rational-expectations-equilibrium diagnostic in public goods and global game settings, the gains from in-context learning vanish when price trends break — though Evolve notes the reasoning null is not airtight.
PersonaPath: a knowledge-centric benchmark for personalized learning paths
Builds an LLM benchmark for personalized learning path planning from 2,000 learner personas and 347 textbooks with knowledge-graph prerequisite relations. The build is rigorous but evaluation is simulation-only, and the best model passes just 29.5% of cases — leaving large headroom.
MIRAGE: conversation state splits evidence-grounding failures into two regimes
A controlled study of multimodal personal agents showing that pre- and post-compaction conversation states produce distinct grounding-failure regimes when using historical evidence. Internally convincing, but six artifacts and one runtime bound how far the finding generalizes.

Timeline

2026First-author arXiv preprint tests whether in-context learning in multi-agent interdependent decision-making reflects recursive reasoning or statistical extrapolation, using a rational-expectations-equilibrium diagnostic.
2026Releases PersonaPath, a knowledge-centric personalized learning path planning benchmark built from 2,000 personas and 347 textbooks.
2026Publishes MIRAGE, a controlled study of how conversation state shapes historical evidence use in multimodal personal agents.

Papers in the Evolve catalog

2026-09-17first author
Recursive Reasoning or Statistical Extrapolation? In-Context Learning in Multi-Agent Interdependent Decision-Making
2026-09-17first author
PersonaPath: Towards Knowledge-Centric Personalized Learning Path Planning
2026-09-17first author
MIRAGE: How Conversation State Shapes Historical Evidence Use in Multimodal Personal Agents

Frequent co-authors

Yu Liu×2 Wenwen Li×1 Yifan Dou×1 Guangnan Ye×1 Zeming Liu×1 Tianle Zhang×1 Zihao Cheng×1 Yuhang Guo×1 Kehai Chen×1 Min Zhang×1 Yunhong Wang×1 Haifeng Wang×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.

Bibliometric dimensions are computed from OpenAlex-indexed records only; no other citation databases are reflected.
No institution, country, or ORCID is available; 'Yu Liu' is a common name, so this profile is name-resolved from public bibliographic data and may merge or split identities.
Output, impact, and activity axes are unscored because the underlying data (publication volume, citations, multi-year cadence) is insufficient.
All catalogued papers are 2026 arXiv preprints; the data contains no peer-reviewed venue, funding, or award information.