Heejin Do

Research profile

Research summary

Heejin Do is a natural language processing researcher at ETH Zurich whose work concentrates on two evenly balanced strands — LLM-based student simulation for educational AI and LLM evaluation (field distribution 0.5/0.5 each). The OpenAlex record lists 20 works in the recent three-year window, with yearly counts rising from 1 (2024) and 2 (2025) to 17 (2026), giving an activity score of 89/100. Topic modeling of this output clusters around topic modeling and NLP techniques (9 each), intelligent tutoring systems and adaptive learning (8), AI in education (7), and online learning and analytics (5). Two Evolve-reviewed 2026 papers anchor the profile. SWIM, first-authored, shows prompting weak, SFT strong, and AES-rewarded GRPO reinforcement learning best for proficiency-conditioned student writing simulation, with high reproducibility readiness. A co-authored study applies Knowledge Space Theory to LLM mathematical reasoning, finding no human-like prerequisite structure (medium readiness). Do also co-authored GPAgentBench-2K, a benchmark for LLM agents in complex clinical action spaces. Early-stage indicators — h-index 1, i10-index 0 — mean output and impact scores are not yet available. Frequent co-authors include Mrinmaya Sachan, Boqi Chen, Xudong Liu, Yunke Ao, Jianing Qiu, Jakub Kontak, and Peng Cui.

Research keywords

Highlights

• SWIM: proficiency-conditioned student writing simulation, best with AES-rewarded RL

  First-authored 2026 study proposing SWIM (Student Writing Simulation via Proficiency-Conditioned Generation). Evolve's verdict: prompting is weak, SFT strong, and AES-rewarded RL (GRPO) best at simulating student essays conditioned on trait scores — credible, but the evaluation rests on an automated essay scoring proxy. Reproducibility readiness: high.

• Do LLM math knowledge structures look human-like? A Knowledge Space Theory test

  Co-authored 2026 arXiv study applying Knowledge Space Theory to LLM mathematical reasoning. Evolve's verdict: this novel KST-based evaluation finds LLM math knowledge lacks human-like prerequisite structure — persuasive, though annotation noise and the absence of significance tests temper it. Reproducibility readiness: medium.

• Co-authorship on GPAgentBench-2K, a clinical LLM-agent benchmark

  Do appears as a middle author on GPAgentBench-2K: Benchmarking Large Language Model Agents in Complex Clinical Action Space (2026, arXiv). This extends the profile beyond education into clinical agent evaluation. Evolve has not yet reviewed it, so no verdict, keywords, or reproducibility rating is on record.

Timeline

1. 2026First-authored SWIM introduces proficiency-conditioned student writing simulation, with prompting weak, SFT strong, and AES-rewarded GRPO reinforcement learning rated best; reproducibility readiness is high.
2. 2026Co-authored the Knowledge Space Theory study finding that LLM mathematical reasoning knowledge lacks human-like prerequisite structure; reproducibility readiness is medium.
3. 2026Contributed as a middle author to GPAgentBench-2K, benchmarking LLM agents in complex clinical action spaces.
4. 2026Indexed output surged to 17 works in 2026, up from 1 in 2024 and 2 in 2025, per OpenAlex countsByYear, with 20 works over the recent three-year window.

Papers in the Evolve catalog

• 2026-09-04

  Do LLMs Exhibit Coherent Knowledge Structures in Mathematical Reasoning? A Perspective from Knowledge Space Theory

• 2026-09-02first author

  SWIM: Student Writing Simulation via Proficiency-Conditioned Generation

Frequent co-authors