Early-career LLM-agent reliability researcher (calibration and memory) at Shaanxi University of Technology
2
papers here
2
reviewed by Evolve
2
first-author
2
corresponding
—
works · OpenAlex
0
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—
This axis measures total publication volume and productivity trajectory, but no aggregated OpenAlex output count is available for this record; evidence is limited to the two 2026 arXiv preprints in the Evolve catalog, so output remains unscored.
Impact—
This axis would measure citation uptake — citations, h-index, i10-index and field-weighted influence. The record lists h-index and i10-index as 0 with no citation streams for these brand-new 2026 preprints, so impact cannot be scored yet.
Focus0/100
The raw field distribution is tightly concentrated — 50% artificial intelligence (LLM agent memory) and 50% LLM agent evaluation and confidence calibration — even though the small corpus yields a nominal focus score of 0. Topic counts agree: software engineering research (2) plus single mentions of testing and debugging, retrieval, planning, and adversarial robustness, sketching a coherent early agenda on LLM agent reliability.
Collaboration0/100
Scores 0 because the record resolves zero distinct coauthors and lists no frequent collaborators, matching the solo, first-and-corresponding authorship of both 2026 preprints. This may partly reflect unresolved author entities rather than genuinely isolated work. A collaboration network cannot yet be characterized and should be re-evaluated as co-authored output accumulates.
Activity32/100
Scores 32 from a very recent but real burst: counts-by-year show 2 works in 2026, matching recent3yWorks of 2 and the two catalog preprints. Both are first-authored and corresponding, indicating substantive individual output rather than incremental contributions. The rhythm is a single 2026 cluster, so the score should be read as early-stage momentum, not sustained multi-year production.
Reproducibility78/100
The strongest axis: mean reproducibility readiness of 78 across 2 scored papers. The calibration-audit preprint rates high, consistent with its frozen, testing-driven audit design over SWE-bench and TerminalBench; the procedural-memory study rates medium, reflecting 32 one-shot qwen3:8b cells with n=1 per cell. The pattern indicates deliberate emphasis on controlled, re-runnable protocols at an early career stage.
Research summary
YanZe Cao is an early-career researcher at Shaanxi University of Technology (China), name-resolved from public OpenAlex bibliographic data. The record holds two 2026 arXiv preprints, each listing Cao as first and corresponding author, together forming a coherent agenda on the reliability of LLM-based agents. The first paper runs a frozen, testing-driven reliability audit of trajectory-based early outcome prediction for LLM agents; it finds that target-specific calibration transfer persists only within a single benchmark (SWE-bench), with no cross-benchmark replication on TerminalBench — a credible but deliberately narrow result. The second studies procedural memory under change in controlled web tasks, reporting that mismatched procedural memory showed no predefined interference across 32 one-shot qwen3:8b cells on WebShop-style tasks; again careful but narrow, with one run per cell. Evolve scores reproducibility readiness at a mean of 78 across the two papers (high and medium), the strongest quantitative signal in this profile. Aggregate indicators remain minimal: counts-by-year show 2 works in 2026, recent-3-year works total 2, and h-index and i10-index are both 0, so output and impact are left unscored. The field distribution splits evenly (50/50) between LLM agent memory and LLM agent evaluation with confidence calibration.
Collaboration pattern
The OpenAlex record resolves zero distinct coauthors and the frequent-coauthor list is empty, so the collaboration dimension scores 0 — likely reflecting solo-authored preprints or unresolved author entities. Both catalog papers position Cao as first and corresponding author, consistent with an independent, self-driven working mode. No collaboration network can be characterized yet; this axis should be revisited as co-authored work and author disambiguation accumulate.
Research keywords
Highlights
Frozen audit shows calibration transfer holds only within one benchmark
Careful null result: mismatched procedural memory showed no predefined interference
Timeline
- 2026First catalog preprint: testing-driven reliability audit of trajectory-based early outcome prediction for LLM agents (first and corresponding author, arXiv).
- 2026Second catalog preprint: procedural memory under change — reuse and interference in controlled web tasks (first and corresponding author, arXiv).
- 2026OpenAlex record shows 2 works in 2026, both in the Evolve catalog, with a reproducibility mean of 78 across the 2 scored papers.
Papers in the Evolve catalog
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.
- All bibliometrics come from an OpenAlex name-resolved record (identity: openalex); no ORCID is linked, so attribution carries normal name-disambiguation risk for this name.
- h-index and i10-index are recorded as 0 and no citation counts are available, so the output and impact dimensions are unscored rather than estimated.
- The reproducibility mean (78) is computed from only 2 reviewed papers, both 2026 arXiv preprints; conclusions may shift as the record grows.
- The collaboration score of 0 reflects zero resolved coauthors and may indicate unresolved author entities rather than confirmed solo authorship.