Xue-Yong Fu

Research profile

Research summary

Xue-Yong Fu is a natural language processing researcher whose profile is name-resolved from OpenAlex bibliographic data; no institution, country, or ORCID is attached to this snapshot. The record shows a compact but active output profile: an h-index of 6, an i10-index of 5, and 15 works in the last three years, with per-year counts peaking at 10 in 2022 and holding at five to six across 2025–2026. Topical mass concentrates in Topic Modeling (26) and Natural Language Processing Techniques (17), with smaller clusters in advanced text analysis, speech recognition and synthesis, and data quality. The two Evolve-reviewed papers, both 2026 arXiv preprints, define a clear current agenda: controlled comparisons of LLM adaptation strategies. One study compares fine-tuning strategies for customer-support LLMs across 13 models and finds multi-task full fine-tuning best; the other runs SFT-versus-GRPO experiments across six Qwen3 scales and three tool-calling corpora, concluding that LoRA-SFT is the strongest in-distribution method and that dataset choice matters as much as method. Evolve credits the controlled designs (reproducibility mean 78) but flags missing significance tests in both, a private-data limitation in one, and an eval-metric-reward alignment that softens one transfer claim. Collaboration is small-scale (5 distinct coauthors), with Shashi Bhushan TN and Md Tahmid Rahman Laskar recurring; the author sits in middle, non-corresponding positions on both cataloged papers.

Research keywords

Highlights

• LoRA-SFT emerges as the strongest in-distribution recipe for tool-calling agents

  A controlled SFT-vs-GRPO-vs-SFT→GRPO study across six Qwen3 scales and three tool-calling corpora concludes LoRA-SFT is the strongest in-distribution method and that dataset choice matters as much as method; Evolve flags the missing significance tests and the eval-metric-equals-reward alignment that softens the transfer claim.

• One adapted model for customer support? A 13-LLM comparison favors MT-full

  A controlled comparison of fine-tuning strategies across 13 customer-support LLMs finds multi-task full fine-tuning (MT-full) best; Evolve finds the result useful but notes the absence of significance tests and that half the data is private.

• A signature method: controlled comparison of adaptation strategies

  Both cataloged 2026 papers share one playbook — isolate data, method, and scale in controlled comparisons of LLM adaptation. Evolve's reproducibility mean for this author is 78 across the two scored papers, though both verdicts call for significance testing.

Timeline

1. 2021Earliest year in the OpenAlex per-year counts, with 2 works recorded.
2. 2022Peak year in the snapshot, with 10 works recorded.
3. 2024Counts bottom out at 4 works for the year (down from 5 in 2023).
4. 2025Output rebounds to 6 recorded works.
5. 20265 works recorded so far, including the two arXiv preprints in the Evolve catalog — the fine-tuning strategy selection study for support LLMs and the SFT-vs-RL tool-calling study.

Papers in the Evolve catalog

• 2026-09-23

  Can One Adapted Model Do It All? Fine-Tuning Strategy Selection for Customer Support LLMs

• 2026-09-18

  When Better Turns Do Not Make Better Agents: Diagnosing the Gap Between Next-Turn Metrics and Workflow Success

• 2026-09-17

  SFT or RL for Tool-Calling Agents? A Controlled Study Across Data, Method, and Scale

Frequent co-authors