Understanding Performance of Long-Document Ranking Models through Comprehensive Evaluation and Leaderboarding

We evaluated 20+ Transformer models for ranking of long documents (including recent LongP models trained with FlashAttention) and compared them with simple FirstP baselines (applying the same model to input truncated to the first 512 tokens). We used MS MARCO Documents v1 as a primary training set and evaluated models in the zero-shot scenario as well as after fine-tuning on other collections. In our initial experiments with standard collections we found that long-document models underperformed FirstP or outperformed it by at most 5% on average in terms of MRR or NDCG. We then conjectured that this was not due to models inability to process long context but rather due to a positional bias of relevant passages, which tended to be among the first 512 document tokens. We found evidence that this bias was, indeed, present in at least two test sets, which motivated us to create a new collection MS MARCO FarRelevant where the relevant passages were not present among the first 512 tokens. Unlike standard collections where we observed both little benefit from incorporating longer contexts and limited variability in model performance (within a few %), experiments on MS MARCO FarRelevant uncovered dramatic differences among models. FirstP models performed roughly at the random-baseline level in both zero-shot and fine-tuning scenarios. Simple aggregation models (e.g., MaxP) had good zero-shot accuracy but benefited little from fine-tuning. Most other models had poor zero-shot performance (sometimes at a random baseline level) but outstripped MaxP by as much 13-28\% after finetuning. Thus, positional bias not only diminishes benefits of processing longer document contexts but also leads to model overfitting to this bias and performing poorly in a zero-shot setting when a distribution of relevant passages changes substantially. We make our software and MS MARCO FarRelevant available.