Tag
The paper introduces SAE-IDF, which uses sparse autoencoders and rarity-weighted features as a shared coordinate system to match symbolic units across heterogeneous systems, outperforming dense LLM baselines on 12 tasks while remaining human-interpretable.
The author details a project on automating book digitization for Urdu literature by recovering crop labels from manual Photoshop work, finding that ten operator clicks per book outperformed scaling data or model complexity, and discusses retouching methods using neural nets and classical techniques.