This article from LightOn AI researchers describes a Straight-Through Estimator-based regularization technique that improves hierarchical pooling in ColBERT-style retrieval models. Combined with adaptive budget allocation, the method achieves 99.4% retention at 5x compression without degrading full-token retrieval quality, building on prior MUVERA regularization work. The team reports a 22.3 percentage-point improvement in retention metrics, arguing that directly training for compressibility outperforms post-hoc compression.
