Youngjae Cho
ML Research Engineer
Seoul, South Korea
I study robust and data-efficient learning from imperfect supervision: preference optimization, active learning, Bayesian adaptation, and reinforcement learning.
Education
KAIST · M.S., Industrial & Systems Engineering
Mar 2022 – Feb 2024KAIST · B.S., Industrial & Systems Engineering
Mar 2017 – Feb 2022Selected Publications
* Equal contribution.
-
Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment
NeurIPS 2026 · Accepted · First author [Paper]
Proposed a batch-conditioned adversarial anchor and anchor-gap reweighting for brittle preferences; improved AlpacaEval 2.0 LC win rate by 3.6 percentage points over SimPO (Mistral-7B).
-
Make Prompts Adaptable: Bayesian Modeling for Vision-Language Prompt Learning with Data-Dependent Prior
AAAI 2024 · First author [Paper][Code]
Developed Bayesian prompt adaptation with data-dependent priors and Wasserstein gradient flow for posterior inference, improving few-shot generalization under distribution shift.
-
Sharpness-Aware Active Learning
ICML 2023 · Co-first author [Paper][Code]
Co-developed sharpness-aware acquisition through pseudo-label-based perturbed-loss estimation, with a theoretical optimization bound; evaluated classification, detection, and domain-adaptive segmentation.
-
Beyond Scalar IoU: Structured Verification from Rollout Groups for Video Temporal Grounding
arXiv 2026 · Preprint · First author [Paper]
Introduced a group-conditioned verifier with an exact decomposition into IoU and a covariance correction; improved temporal grounding at every reported IoU threshold across five benchmarks.
Research Experience
Pyler
Oct 2025 – Present- Preference learning. First-author GAPO research on local margin degradation under parameter perturbations; evaluated robustness to random and structured preference noise.
- Structured verification. First-author SUTURE research on group-conditioned rewards; analyzed verifier structure and trained temporal-grounding policies with GRPO.
- Multimodal post-training. Implemented rollout, preference loss, and training loop on Megatron-Bridge for Nemotron-Nano-12B-VL; designed SimPO preference data for video localization and content safety.
- Auditable supervision. Recovered decision trees from 299K reasoning traces; improved content-safety macro F1 from 0.777 to 0.857.
Aiv Co.
Mar 2024 – Oct 2025- Generative modeling. Led background-aware diffusion research for scarce industrial anomalies; evaluated defect synthesis on MVTec-AD and LOCO and deployed it to improve production detector precision and recall.
KAIST
Mar 2022 – Feb 2024- Statistical learning. First-author APP and co-first-author SAAL research combining Bayesian inference, optimization, and active learning; evaluated adaptation under distribution shift and acquisition across classification, detection, and segmentation.
Additional Publications
-
Background-Aware Defect Generation for Robust Industrial Anomaly Detection
arXiv 2024 · Preprint · First author [Paper]
-
Improving Group-based Robustness and Calibration via Ordered Risk and Confidence Regularization
ICML 2022 Workshop [Paper]
Workshop on Spurious Correlations, Invariance and Stability
Academic Service
- ICML 2026
- Gold Reviewer
- NeurIPS 2026
- Top Reviewer
Awards
NVIDIA Nemotron Hackathon
2026Team winner in Track B (Domain-Specialized Model). Implemented the preference-optimization post-training (SimPO on Megatron-Bridge) behind the team's winning video content-safety VLM.