AI Dynamics

Global AI News Aggregator

About

Position Bias in LLMs: How Architecture and Training Data Matter

LLMs sometimes exhibit position bias, where they overemphasize the beginning or end of a document or conversation while neglecting the middle. A theoretical framework from MIT revealed that model architectures & training data contribute to this problem: http://
bit.ly/44mW9V8

→ View original post on X — @mit_csail