Industrial retrieval systems often chase better quality by expanding embedding candidate pools or adopting complex training pipelines. A new arXiv paper, Beyond Recall Saturation: Scaling Hard-Negative Reranker Training for Industrial Legal Retrieval, challenges that approach. The authors show that under a sufficiently large but practical candidate pool, embedding recall exhibits saturation — meaning further enlarging the pool yields diminishing returns.
Instead, the paper argues that the bottleneck shifts to the reranker stage, specifically the training of hard negatives. The authors suggest that scaling hard-negative reranker training is a more effective way to improve retrieval quality in industrial legal settings than simply growing the candidate pool.
The finding has practical implications for system design: rather than investing in ever-larger embedding recall, engineers may need to allocate resources to reranker training data and strategies. The paper is available on arXiv under identifier 2610.24478.