Eugene Wen

—

Unclaimed · AI-structured profileOpenAlex

Eugene Wen — LLM evaluation researcher probing long-horizon procedural reasoning, agent memory, and prompt stability

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—

Aggregate output metrics are not yet available in Evolve's snapshot; this axis would measure total publication volume and its cadence over time. Only per-year counts and 7 recent three-year works are on file, so no score is assigned.

Impact—

Citation impact data is insufficient in the current snapshot; this axis would measure citation accumulation, per-paper influence, and scholarly uptake over time. With no citation totals on file, the axis is left unscored rather than estimated.

Focus0/100

Focus scores 0, reflecting a strongly two-phase record rather than a single line. The OpenAlex topic profile splits across five clusters — Healthcare Quality and Management (8), Medical Coding and Health Information (6), Patient Safety and Medication Errors (5), Topic Modeling (4), and Natural Language Processing Techniques (2) — while the recent field distribution is split evenly (0.5/0.5) between LLM agents and LLM evaluation and benchmarks. A 2011–2024 publishing gap separates the two phases.

Collaboration44/100

Collaboration scores 44, a moderate level: the recent three-year window contains 7 works with 8 distinct coauthors. Repeated collaborators include Syed Shariyar Murtaza, Utkarsh Soni, Yifan Nie, Wentao Zhang, Junaid Ahmad Bhatti, Yuntian Deng, Arvid Frydenlund, and Sachin Chandrasekhar. All three 2026 catalog papers are multi-authored, with Wen as last author once and in middle positions twice, consistent with team-based LLM research.

Activity61/100

Activity scores 61, driven by an accelerating recent cadence: per-year counts show 1 work in 2024, 1 in 2025, and 5 in 2026, totaling 7 works in the last three years. This follows a long quiet stretch — the record shows earlier output from 2001 through 2011, peaking at 6 works in 2008, before a gap until 2024 — making the current cadence the strongest of the recorded career.

Reproducibility55/100

Reproducibility scores 55, computed from Evolve's readiness assessment of 2 catalog papers (mean 55). Readiness varies across the three reviewed 2026 papers: 'Control-Data Flow Separation' rates high, while the TAM benchmark paper and the synthetic-data forgetting study each rate medium; the TAM verdict also flags single-model evaluation and an unquantified label-noise ceiling as limitations.

Research summary

Eugene Wen is a researcher whose recent work centers on evaluating and stabilizing large language model systems: benchmarks for long-horizon procedural reasoning, continual-learning effects in LLM-agent skill retrieval, and prompt optimization for multi-agent systems. Evolve's catalog includes three 2026 arXiv papers. As last author on 'Tasks over Application Manuals,' Wen contributed to the TAM benchmark, which shows GPT-5 near-failing long manual procedures with 1%–15.5% exact match. 'When Synthetic Data Hurts' shows synthetic-data LoRA tuning causing catastrophic forgetting in skill retrieval, with low rank outweighing regularizers, while 'Control-Data Flow Separation' achieves 100% output validity for multi-agent prompt optimization by construction. The OpenAlex-linked bibliometric record is distinctive: counts by year show earlier output from 2001–2011 concentrated in healthcare topics — Healthcare Quality and Management (8), Medical Coding and Health Information (6), Patient Safety and Medication Errors (5), plus Topic Modeling (4) and Natural Language Processing Techniques (2) — followed by a gap and a 2024–2026 return totaling 7 recent three-year works. The recent field distribution splits evenly between LLM agents and LLM evaluation and benchmarks. An h-index and i10-index of 7 reflect a compact record. Because the name is common, this profile is name-resolved from public bibliographic data.

Collaboration pattern

Wen's recent work is team-based: 7 works in the last three years with 8 distinct coauthors, and all three 2026 catalog papers are multi-authored, with Wen as last author once and in middle positions twice. Recurring collaborators include Syed Shariyar Murtaza, Utkarsh Soni, Yifan Nie, Wentao Zhang, Junaid Ahmad Bhatti, Yuntian Deng, Arvid Frydenlund, and Sachin Chandrasekhar. The moderate collaboration score (44) suggests a small, stable core group rather than a large rotating network.

Research keywords

LLM evaluationBenchmark constructionLong-horizon procedural reasoningLLM agentsSkill retrievalCatastrophic forgettingSynthetic dataContinual learningLoRA fine-tuningPrompt optimizationMulti-agent LLM systemsRetrieval-augmented generation

Highlights

TAM benchmark exposes near-failure of GPT-5 on long manual procedures
As last author, Wen contributed to 'Tasks over Application Manuals,' a benchmark of long application manuals where GPT-5 achieves only 1%–15.5% exact match, exposing gaps in long-horizon procedural reasoning for language models; Evolve values the contribution while noting single-model evaluation and an unquantified label-noise ceiling.
Showing when synthetic data hurts LLM-agent skill retrieval
In 'When Synthetic Data Hurts' (middle author), synthetic-data LoRA fine-tuning causes catastrophic forgetting in skill retrieval for LLM agents; regularized tuning preserves out-of-distribution performance, but low rank outweighs the regularizers — a practical warning for continual-learning pipelines.
Control-data flow separation stabilizes multi-agent prompt optimization
In 'Control-Data Flow Separation' (middle author), frozen control channels keep prompt optimization from corrupting multi-agent protocols, achieving 100% output validity by construction; Evolve rates its reproducibility readiness high while noting that quality gains are modest and untested.

Timeline

2001Earliest recorded publication year in the OpenAlex counts-by-year record (1 work).
2008Peak of the early healthcare-focused publishing phase, with 6 works recorded.
2011Last year of the early phase before a publishing gap (1 work recorded).
2024Return to publishing after the gap: 1 work recorded, opening the current LLM-focused phase.
2026Most productive recorded year: 5 works, including all three Evolve catalog papers — the TAM benchmark, the synthetic-data forgetting study, and control-data flow separation for multi-agent prompt optimization.

Papers in the Evolve catalog

Frequent co-authors

Syed Shariyar Murtaza×3 Utkarsh Soni×3 Yifan Nie×3 Wentao Zhang×1 Junaid Ahmad Bhatti×1 Yuntian Deng×1 Arvid Frydenlund×1 Sachin Chandrasekhar×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 figures (h-index 7, i10-index 7, counts by year, topics) come from an OpenAlex-linked snapshot; no institution, country, or ORCID was provided, and the name is common — this profile is name-resolved from public bibliographic data.
Output and impact dimensions are unscored because aggregate citation and output metrics are not yet available in the snapshot.
The 2012–2023 publishing gap is inferred from missing counts-by-year entries; no career or biographical information is available to explain it.
Qualitative verdicts and reproducibility-readiness ratings reflect Evolve's review of only three 2026 arXiv papers; the reproducibility score (55) is computed from 2 of them.