Daeun Ji

Research profile

Research summary

Daeun Ji is an early-career machine-learning researcher at Yonsei University (South Korea), whose publicly indexed record — resolved via ORCID 0009-0006-9507-4580 and OpenAlex — is concentrated almost entirely on time series forecasting with large language models. The field distribution splits evenly (0.5 / 0.5) between LLM-based time series forecasting and general time series forecasting, and OpenAlex topics mirror this: Forecasting Techniques and Applications (3) and Time Series Analysis and Forecasting (3) lead, with Stock Market Forecasting Methods (2) and single works on topic modeling and sensor/energy-harvesting materials. The Evolve catalog includes two 2026 arXiv papers. As first author of "Beyond Similarity: Coverage-Aware Prompt Selection for Time Series Forecasting with LLMs", Ji proposes a parameter-free coverage regularizer for similarity-based prompt retrieval; Evolve notes unusually honest statistical reporting but an aggregate gain that resolves on only 1 of 24 cells. The middle-authored "CTRL" casts LLM agents as label-free residual controllers for frozen forecasting backbones, though missing error bars and a Table 3 mismatch weaken its rigor. Bibliometric signals are those of an emerging profile: h-index 1, i10-index 1, 5 works in the last three years, and 6 distinct co-authors. Radar scores stand at 78 for reproducibility, 52 for activity, 39 for collaboration, and 0 for focus — the latter reflecting a deliberately narrow specialization; output and impact are not yet scored.

Research keywords

Highlights

• First-authored coverage-aware prompt selection for LLM forecasting

  On "Beyond Similarity: Coverage-Aware Prompt Selection for Time Series Forecasting with LLMs" (arXiv, 2026), Ji is first author, introducing a parameter-free coverage regularizer that counters similarity-based prompt selection bias in LLM-based forecasting. Evolve credits unusually honest statistics, while noting the aggregate gain resolves on only 1 of 24 cells; reproducibility readiness is rated high.

• LLM agents as label-free residual controllers for frozen forecasters

  "CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning" (arXiv, 2026, middle author) casts LLM agents as label-free residual controllers over frozen time-series-forecasting backbones, combining residual learning, test-time adaptation and STL decomposition to handle distribution shift. Evolve notes the gains track shift, but absent error bars and a Table 3 mismatch weaken rigor; reproducibility readiness is medium.

• Reproducibility is the profile's strongest radar axis (78)

  Across the 2 papers Evolve has assessed, the mean reproducibility readiness is 78 — well above this profile's other scored axes (activity 52, collaboration 39, focus 0), driven by high readiness on "Beyond Similarity" and medium on "CTRL".

Timeline

1. 2024Earliest indexed output: OpenAlex countsByYear records 2 works in 2024, the first year with publication activity in the snapshot.
2. 2026Most active year: 3 works recorded in 2026, the peak of the countsByYear series (2024: 2; 2026: 3).
3. 2026Posts "CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning" to arXiv as a middle author; Evolve rates its reproducibility readiness medium.
4. 2026Posts "Beyond Similarity: Coverage-Aware Prompt Selection for Time Series Forecasting with LLMs" to arXiv as first author; Evolve rates its reproducibility readiness high.

Papers in the Evolve catalog

• 2026-09-20

  CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning

• 2026-09-19first author

  Beyond Similarity: Coverage-Aware Prompt Selection for Time Series Forecasting with LLMs

Frequent co-authors