Paper Summaries
This paper addresses the performance gap in fine-tuned LLMs caused by distribution mismatch between the model and task dataset, proposing a self-distillation approach to bridge this gap.
SelectIT introduces a novel approach to instruction tuning that uses uncertainty-aware self-reflection to selectively choose high-quality training examples for LLMs.
End-to-end sparse autoencoder approach for learning interpretable representations in neural networks.
Exploring how CNN architecture itself holds powerful priors for image restoration tasks like super-resolution, denoising, and inpainting - without any training data.