Article

Anchored Answers: Unravelling Positional Bias in GPT-2's Multiple-Choice Questions
Using principles from Mechanistic Interpretability, we are able to investigate and correct the positional bias in GPT2 models regarding multiple-choice questions.
PDF Dataset Poster Video Source Document
Learning to Maximize Mutual Information for Chain-of-Thought Distillation
We propose a new training paradigm for Chain-of-Thought distillation with the perspective from information bottleneck.
PDF Code Dataset Poster Video Source Document
Learning to Maximize Mutual Information for Chain-of-Thought Distillation