ConceptTargeted Preference Optimization
Targeted Preference Optimization
A technique that restricts preference optimization to already identified key circuits, so that updating only a small number of parameters effectively improves a model's performance on specific tasks.
Timeline (last 90 days)
Oct 16
The core idea of the TITAN framework is to shift the optimization of translation capability from full-parameter fine-tuning to causal circuit localization and targeted optimization, integrating three technical paths: causal intervention, sparse feature decomposition, and targeted preference optimization.
Unverified50%