Summary

PListMLELoss (and ListMLELoss, which subclasses it) scatter a list's logits into a
(batch_size, max_docs) matrix and pad shorter lists with a logit of 1e-16. The Plackett-Luce
log-likelihood uses a reverse cumulative sum of scores = logits.exp() as the per-position
normalizer. Because exp(1e-16) ~= 1, every padded slot enters that normalizer with roughly unit
mass, inflating the denominator of every real document's log-probability. The padding mask is only
applied after the cumulative sum (log_probs[~mask] = 0.0), so it removes the padded positions'
own terms but not their pollution of the real positions' normalizers.

As a result, a query's loss and gradients depend on max_docs — i.e. on how many documents the
longest other query in the batch happens to have. For a single list (real labels [2, 1],
logits [1.5, 0.3]):

The same list's loss grows with padding width and the second document's gradient flips sign, so a
variable-length-list batch (the common case for reranker training) trains on a corrupted signal.

This excludes padded entries from the normalizer by zeroing their scores before the cumulative sum,
using a mask in the same (sorted) order as the logits:

scores = sorted_logits.exp().masked_fill(~sorted_mask, 0.0)

This matches how reference ListMLE implementations handle padding (e.g. allRank sets padded
predictions to -inf before the exp, which is the same as zero mass in the normalizer). For a list
with no padding the change is a no-op, so existing single-length behavior is unchanged.

Tests

Added a deterministic padding-invariance regression test for ListMLELoss and PListMLELoss,
parametrized over respect_input_order (True/False): it batches a 2-document query with a
3-document query and asserts the batched loss equals the mean of the two per-query losses. It fails
on current main and passes with this change:

pytest tests/cross_encoder/losses/test_listwise_padding.py -q
# main:  4 failed
# fix:   4 passed