Dongyoon Han

Research profile

Research summary

Dongyoon Han is a researcher at NAVER Cloud (South Korea) whose recent work centers on post-training and interpreting large language models. The OpenAlex profile records 28 works in the last three years, with output accelerating sharply to 25 works in 2026 and 14 distinct co-authors; citation-based indicators remain early-stage (h-index 2, i10-index 1), typical for such a recent burst. Three strands recur across the Evolve-reviewed catalog. First, Mixture-of-Experts post-training: Router Prior Bias (2026) argues that soft-anchoring routers preserves base routing structure better than re-applied load-balancing losses, with interventional evidence that Evolve rates careful but Moonlight-centric. Second, reasoning-trace curation: Revisiting Complete Reasoning Traces for Post-Training (2026) proposes trimming the middle of traces (E-SFT), with gains across SFT, RL, and distillation that are consistent but small. Third, mechanistic interpretability: Beneath the Surface of Chains-of-Thought (2026) finds reasoning operations form separable middle-layer geometry. A fourth 2026 paper, not yet reviewed, addresses verification-aware training for speculative decoding. Earlier OpenAlex topics — topic modeling (18), advanced neural network applications (12), domain adaptation and few-shot learning (12) — indicate a broader machine-learning base preceding the current LLM focus.

Research keywords

Highlights

• Router Prior Bias: anchoring MoE routers in post-training

  Evolve's verdict: soft-anchoring MoE routers preserves base routing structure better than re-applied load-balancing losses in post-training; the interventional evidence (expert co-activation graphs, NMI community structure) is careful, but the study is Moonlight-centric. Repro-readiness: medium.

• E-SFT: trimming the middle of reasoning traces

  Evolve's verdict: trimming the middle of complete reasoning traces (E-SFT) improves SFT, RL, and distillation for large reasoning models; gains are consistent but small and lack statistical rigor. Rated high repro-readiness — the most replication-ready paper in the catalog.

• Mechanistic geometry of chain-of-thought reasoning operations

  Evolve's verdict: chain-of-thought reasoning operations form separable middle-layer geometry, well controlled for confounds via linear probing and attention masking; GPT-5 label noise and split leakage temper the mechanistic claims. Repro-readiness: medium.

• Verification-aware training for speculative decoding

  A fourth 2026 catalog paper in which Han holds the last author position. It has not yet been reviewed by Evolve — no verdict, keywords, or repro-readiness rating is available, so it is listed here for completeness only.

Timeline

1. 2021Earliest recorded OpenAlex work in countsByYear (1 work), preceding the current LLM-focused phase.
2. 20242 works recorded — a quiet period before the 2026 surge.
3. 20251 work recorded in OpenAlex countsByYear.
4. 2026Breakthrough year: 25 works in countsByYear and all four catalog papers — Router Prior Bias (MoE post-training), E-SFT reasoning-trace trimming, the mechanistic chain-of-thought study, and verification-aware training for speculative decoding.

Papers in the Evolve catalog

• 2026-09-08

  Router Prior Bias: Preserving Base Routing Structure in MoE Post-Training

• 2026-09-07

  Revisiting Complete Reasoning Traces for Post-Training

• 2026-09-04

  Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs

Frequent co-authors