A method for calculating positional bias uses the formula: $f \frac{\psi(i - j)}{\psi(\max(\text{mlen}, i))}$ where i is the current token's position, j is another token's position, mlen is a fixed length parameter, f is a scaling factor, and ψ is a monotonically increasing function (i.e., its value increases as its input increases). How does the normalization term in the denominator, ψ(max(mlen, i)), affect the calculated bias values as the current position i grows significantly larger than mlen?

Positional Bias Calculation

Analyzing Positional Bias Behavior