Four papers, one question: how do you train well when the data is too little or the labels are wrong? Each pulls a different lever — which samples to label, how to adapt when there are few, how to synthesize what's missing, and how to keep the optimization stable when the signal is noisy. First or co-first author on all four; two published at ICML 2023 and AAAI 2024, two preprints.
Ref 01/04arXiv 2026Preprint★ First author
GAPO — Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment
Youngjae Cho, Jongsuk Kim, Ji-Hoon Kim
Stability · noisy preferences
Reads preference optimization as learning dynamics, and puts the stability problem in the reference: DPO's frozen anchor becomes a geometrically perturbed one — a worst-case local perturbation of the current policy — with each pair reweighted by its geometric brittleness. A steadier objective, which is what holds up under noisy labels and on less data.
AlpacaEval 2.0 LC +3.6pp vs SimPObeats DPO · SimPO · KTO · ORPO
Background-Aware Defect Generation for Robust Industrial Anomaly Detection
Youngjae Cho, Gwangyeol Kim, Sirojbek Safarov, Seongdeok Bang, Jaewoo Park
Synthesis · scarce anomalies
When real defects are too rare to train on, generate them — diffusion synthesis that models the relationship between foreground defect and background surface explicitly, and argues the disentanglement theoretically rather than only showing it. The method behind the production system on the product page.
Bayesian, data-dependent priors that let vision-language prompts adapt per input instead of staying fixed — so a VLM holds up when there is little data to tune on, and stays calibrated when the input shifts.
An acquisition function that bridges sharpness-aware minimization and active learning — select the samples that flatten the loss landscape, and a fixed labeling budget buys more generalization per label.
Neither is public and neither has a result to report. They're here because they're where the work is going, and because both extend threads already on this site — the temporal localization behind the hackathon entry, and the safety side of the moderation stack.
UnpublishedNo result yet
Video temporal grounding
Locating the moment inside a long video that a description actually refers to.
Findings withheld
UnpublishedNo result yet
Safe diffusion
Keeping a diffusion model from generating what it shouldn't.