Multi-Budget Training Makes Retrieval Model Compression More General

antoine_chaffin · x · 2026-07-07

Instead of training for a single budget, the author trained the model with multiple compression budgets ([4,8,16,32,64,128,300]) at each training step. This multi-budget strategy makes the model easier to pool at any compression level. However, if the deployment target is known, training specifically for r=32 can retain 1.6% more information (reaching 97.4% compared to 95.8% for multi-budget) under the same λ, without sacrificing full token quality.

Related event: hpool Enables Flexible and Adaptive Compression for Retrieval Models(5 posts)→

Original post →

More from Research

Research channel →