ml · 2026
epiplexity-dllm
pythonpytorchtransformersdllm
using epiplexity — a measure of how much the model has learned — to guide a diffusion llm toward better generalisation. early experimentation.
end-to-end nlp pipeline on sec 10-q filings predicting abnormal returns. 51.7% directional accuracy vs 46.3% baseline; walk-forward cv to prevent leakage.
fraud signal detection across 20k+ enron emails. balanced random forest on network + sentiment features; auc-roc 0.884 at 2257:1 class imbalance.
structured lottie json animation generation via llms. 60% renderable output after augmented multi-task training across 4k+ animation samples.
browser extension upscaling low-res images on-device. cut avg page load by 7 s. published at acm doi:10.1145/3715335.3735476.