From software engineering to LLM systems: benchmarks, agent skill retrieval, and stable multi-agent prompt optimization
3
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
11
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: total publication counts and venue-normalized productivity could not be computed from the current record. Beyond the 7 works logged in the last three years, this axis would measure cumulative, venue-weighted research output over the full career.
Impact—
Impact scoring is pending: the citation distribution behind this record is incomplete, so citation totals and per-paper influence cannot be responsibly scored. This axis would measure citations, citation velocity, and uptake of the reviewed 2026 preprints once citation data matures.
Focus0/100
Focus scores 0, reflecting breadth rather than concentration: the recent field distribution splits exactly evenly — 0.5 LLM agents/IR and 0.5 LLM evaluation and benchmarks — while bibliographic topics span five clusters, led by Software Engineering Research (18), Software System Performance and Reliability (13), Testing and Debugging Techniques (10), plus Topic Modeling (9). The record reads as a deliberate pivot, not a narrow specialty.
Collaboration44/100
Collaboration scores 44, a moderate pattern. The recent record lists 8 distinct coauthors, with 8 frequent collaborators including Utkarsh Soni, Yifan Nie and Eugene Wen. Across the three Evolve-reviewed papers, authorship is first on one and middle on two, never corresponding — consistent with a hands-on contributor working inside team-led LLM-systems projects rather than a solo line of work.
Activity61/100
Activity scores 61. Recent output is real but uneven: 7 works in the last three years, with per-year counts of 1 in 2023, 3 in 2025 and 4 in 2026 — an accelerating 2025–2026 ramp. Earlier countsByYear data shows sustained production from 2005 through 2019, peaking at 5 works in 2015, before a gap from 2020 to 2022.
Reproducibility55/100
Reproducibility averages 55 (reproMean) across 2 assessed papers. In the catalog, Control-Data Flow Separation is rated high readiness, while the TAM benchmark study and the synthetic-data forgetting study are each rated medium — the latter flagged for single-model setups and an unquantified label-noise ceiling. Artifacts look partly reusable; exact release detail is not recorded here.
Research summary
Syed Shariyar Murtaza (ORCID 0000-0003-3330-4783) is a researcher whose record, resolved from OpenAlex bibliographic data, spans two distinct phases. The bibliometric base is software engineering: among the record's top topics are Software Engineering Research (18 works), Software System Performance and Reliability (13), Software Testing and Debugging Techniques (10), Topic Modeling (9), and Software Engineering Techniques and Practices (8), with publication activity stretching back to 2005. Since 2026, the profile has shifted decisively toward large language model systems: the recent field split is exactly even between NLP/IR for LLM agents and NLP for LLM evaluation and benchmarks, and all three Evolve-reviewed papers are 2026 arXiv preprints in this space. These include a first-author study showing that synthetic-data LoRA tuning causes catastrophic forgetting in LLM-agent skill retrieval; a benchmark (TAM) on which GPT-5 achieves only 1%–15.5% exact match on long-horizon procedural reasoning over application manuals; and a control-data flow separation scheme that keeps prompt optimization from corrupting multi-agent protocols (100% validity by construction). Aggregate indicators are moderate: h-index 11, i10-index 15, 7 works in the last three years, and 8 distinct recent coauthors. Output and impact axes remain unscored pending fuller citation data, and institution and country are not part of this record.
Collaboration pattern
The record shows a moderate collaboration footprint: 8 distinct recent coauthors and a stable set of 8 frequent collaborators, including Utkarsh Soni, Yifan Nie, Eugene Wen, Wentao Zhang, Junaid Ahmad Bhatti, Yuntian Deng, Arvid Frydenlund and Sachin Chandrasekhar. The 2026 arXiv papers place this researcher as first author once and middle author twice, never corresponding — suggesting hands-on technical contribution within team-led LLM-systems groups rather than an independent lab-lead role so far.
Research keywords
Highlights
TAM benchmark: GPT-5 near-fails long manual procedures
First-author: synthetic data can hurt LLM-agent skill retrieval
Control-data flow separation keeps multi-agent prompt optimization stable
Timeline
- 2005Earliest publication activity in the OpenAlex record begins, with 3 works in 2005 and 2 in 2006 — the start of the software engineering phase later reflected in the topic profile.
- 2015Busiest year in the record: 5 publications, the peak of a sustained 2005–2019 run dominated by software engineering, testing and topic modeling topics.
- 2023After 3 works in 2019, output thins to a single 2023 publication following the 2020–2022 gap, before the current restart.
- 2025Per-year counts rise to 3 works in 2025, the ramp toward the current large language model systems phase.
- 2026Three arXiv preprints enter the Evolve catalog — a first-author study on catastrophic forgetting in LLM-agent skill retrieval, the TAM long-horizon procedural reasoning benchmark, and control-data flow separation for multi-agent prompt optimization — marking the pivot to LLM systems.
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 indicators are computed from OpenAlex records only, with no Scopus, Web of Science or Google Scholar cross-check.
- Output and impact radar axes are unscored because the underlying citation distributions are insufficient in the current snapshot; h-index 11 and i10-index 15 are quoted as-is.
- Institution, country and career history are absent from the record; identity is resolved from public bibliographic data via ORCID 0000-0003-3330-4783.
- Only the three 2026 arXiv preprints carry Evolve review verdicts; the software engineering era is represented solely through aggregate topic and per-year counts.