Eduin E. Hernandez

Research profile

Research summary

Eduin E. Hernandez is a researcher at National Yang Ming Chiao Tung University (Taiwan) working in natural language processing, with an OpenAlex field split of exactly 0.5/0.5 between knowledge graphs and KGQA/LLM reasoning. The snapshot, resolved from OpenAlex and linked to ORCID 0009-0008-2812-5311, records an h-index of 3, an i10-index of 1, yearly output of 4 (2021), 3 (2022), 1 (2023), 1 (2024) and 2 (2026), 3 works in the last three years, and 4 distinct coauthors. The topic profile is broad — stochastic gradient optimization techniques lead with 6 works, followed by topic modeling, advanced neural network applications, semantic web and ontologies, and robot manipulation and learning — but the 2026 output converges sharply on knowledge-graph question answering. Two first-author arXiv papers anchor this agenda: THESEUS frames KGQA as question-conditioned graph navigation with gold paths and fidelity metrics, and a controlled study shows that small language models' navigation fidelity and answer accuracy can rank models differently. Evolve's reviews credit both with controlled, metric-explicit design while flagging small benchmarks, no significance tests, and excluded closest baselines. Reproducibility is the strongest measured dimension (mean 78 over 2 reviewed papers); output and impact are not yet scored. This is an early-stage, method-driven profile whose distinctive contribution is making multi-hop graph reasoning traceable rather than answer-only.

Research keywords

Highlights

• THESEUS: KGQA as question-conditioned, traceable graph navigation

  The paper reformulates knowledge-graph question answering as navigation conditioned on the question, trained with reinforcement-learning agents and evaluated against gold paths with explicit fidelity metrics. Evolve calls the benchmark solid while noting that the closest prior baselines were excluded; its reproducibility readiness is rated high.

• Path fidelity can re-rank small language models in KGQA

  A controlled study of small language models finds that navigation fidelity and answer accuracy can rank the same models differently, arguing that evaluation should go beyond final answers with metrics such as Path Edit Distance. Evolve rates the design sound but flags small benchmarks and no significance tests; reproducibility readiness is medium.

• Gold-path benchmarks for multi-hop graph reasoning

  Together, the two 2026 first-author papers build a measurement toolkit for multi-hop navigation: gold paths that make reasoning steps inspectable, plus fidelity metrics that score how faithfully a model traverses the graph. This traceability lens is the profile's distinctive methodological signature.

Timeline

1. 2021Earliest recorded output in the snapshot: 4 works, the highest yearly count on record and predating the current KGQA focus.
2. 20223 works recorded; output holds at its early peak level before the mid-period slowdown.
3. 2023Output slows to 1 recorded work, beginning an intermittent publishing phase.
4. 2024Again 1 recorded work; the quiet stretch before the 2026 return to publishing.
5. 2026Two first-author arXiv papers appear — THESEUS and the small-language-model KGQA evaluation — marking a sharp convergence on knowledge-graph question answering.

Papers in the Evolve catalog

• 2026-09-23first author

  The Path Matters: Evaluating Small Language Models Beyond Answer Accuracy in KGQA

• 2026-09-13first author

  Theseus in the Graph: Towards Traceable Multi-Hop Graph Navigation

Frequent co-authors