Jae Hoon Kim
I work on machine learning for physical and embodied systems — representation learning, optimization, and the structure of models that have to interact with the world.
Recent threads: physics-informed ML, vision & language representation, and sensing on the body.
Also drawn to economics, human behavior, and social-interaction analysis — the same computational lens, turned toward people.
Working papers
- ms. in prep First-Class Per-Site Importance in Materials: Attention-MIL on a Frozen Pretrained BackboneManuscript in prep · target ICML 2027
Three-regime wedge characterizing when attention-MIL beats mean-pool vs supervised oracle; literature-curated per-site importance eval with top-3 recall 71.4% vs random null 10%; AOPC faithfulness win over post-hoc attribution comparators.
- draft Audit-Trail Identity: Identity Economics Under Persistent SensingWorking paper · target JEBO
Two-period self-signaling model in which persistent passive sensors shift identity utility from narrative-compliance to evidentiary consistency. Three propositions plus a headline corollary on the cross-domain heterogeneity that single-channel accounts can't predict.
- pre-registered TempoSurfViT: Surface-Aligned, Temporally-Aware Transformers for Realized Volatility ForecastingPre-registered draft · real-data eval pending
Masked-autoencoder-pretrained ViT that tokenizes IV surfaces along financial axes and consumes a stack of past surfaces to produce calibrated distributional forecasts of realized volatility. Pre-registration committed; real OptionMetrics evaluation pending data access.