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
Position Bias in LLMs: How Architecture and Training Data Matter
By
–
