Part 4 of 12 · 2 min read
A judge's position and length biases are measurable, not just suspected
A judge that quietly prefers whichever answer comes first, or whichever answer is longer, is making a decision based on something other than quality.
Think of a job interviewer who unconsciously rates the first candidate of the day more favorably, simply because that candidate set the bar. This happens regardless of who is actually more qualified. The bias is real and measurable, even if the interviewer would deny having it.
In plain words: self-bias is favouring answers from the same model, verbosity bias is favouring longer answers, and position bias is favouring whichever option is shown first (or last) -- and because a pairwise judge's choice can be simulated as a mix of a genuine quality signal plus these specific extra probabilities, the exact cost of each bias on accuracy, and the exact fix for position bias, can both be measured directly.
A worked example. A judge with no bias correctly picks the better of two answers 85% of the time. Add a 30% position bias (an extra chance of just picking whichever answer is shown first), and accuracy drops to 75%. Asking the judge twice, with the two answers swapped in order, and keeping only the cases where both answers agree, pushes accuracy on the decided pairs up to 92% -- but 42% of all pairs become "undecided" and have to be dropped or re-judged, because the two orderings disagreed. Length bias is more dangerous because it hides: with a 30% length bias, the judge looks great (90% accuracy) when the longer answer happens to be the better one, and terrible (61%) when the longer answer happens to be worse -- the same bias, radically different apparent performance, purely depending on what the calibration set happened to contain.
The key lines. From code/mini-judge/judge.py:
# from code/mini-judge/judge.py
if not swap:
right += judge(True) == better_is_a
else:
r1 = judge(True)
r2 = judge(False) # second order: the first slot now holds B, so a position bias favours B
if r1 == r2:
right += r1 == better_is_a
else:
ties += 1
Read it as: the swap mode calls judge twice, once with each answer in the first slot, and only counts a verdict when both calls agree. A position bias pushes judge(True) toward A and judge(False) toward B. So a biased judge's two calls disagree more often. That is why "swap and require agreement" works as a direct countermeasure for the specific bias being corrected.
Self-bias, verbosity bias, position bias
Self-bias is favouring answers from the same model. Verbosity bias is favouring longer answers. Position bias is favouring the first (or last) option shown.
Example: the 90%-versus-61% accuracy swing from the same 30% length bias, depending only on whether length happened to correlate with quality in that particular batch of pairs, is exactly why a calibration set has to be built so that length and quality are deliberately uncorrelated -- otherwise the bias can look like a strength purely by accident.