Pilsung Kang

Nearly a third of variable-wise gradients directly oppose each other during standard multivariate forecasting, but resolving these conflicts via single-pass gradient surgery reliably boosts accuracy without multiple backward passes.

LLMs can pinpoint past events but fail to keep a coherent and up-to-date record of customer life changes over time.

Nearly all tested LLMs exhibit saturation in question-order effects, undermining the reliability of standard next-token log-probabilities for auditing.

Active learning can boost unsupervised anomaly detection performance by over 12% in complex time series data, tackling the challenge of subtle anomalies head-on.

DeCoFlow achieves a remarkable 98.40% AUROC in continual anomaly detection while ensuring zero parameter forgetting, revolutionizing how we handle sequential data in industrial applications.

Identifying cliff tokens reveals that a single token can drastically shift LLM performance, with targeted removal leading to near-perfect accuracy in mathematical reasoning tasks.

Spot anomalies before they happen: FATE uses disagreement among time-series forecasting models to provide early warning signals without needing labeled anomaly data.