Is Position Bias in Dense Retrievers Built In, or Learned from Data?

Researchers found that eight pretrained dense retrievers pick up positional preferences from synthetic training data, favoring evidence placed at the start, middle, or end of documents in ways that mirror the training distribution.

When the training data was balanced across document positions, positional sensitivity dropped by 57% to 87% on position-aware benchmarks.

The balanced approach kept retrieval performance competitive, suggesting a practical way to reduce bias without sacrificing quality.