Improving Recursive Transformers with Mixture of LoRAs

TL;DR

This paper introduces Mixture of LoRAs (MoL), a lightweight mechanism that enhances recursive transformers' expressivity through token-conditional weight modulation, achieving state-of-the-art results in compact models.

Contribution

The paper presents MoL, a novel conditional-computation approach that improves parameter sharing in recursive transformers without losing expressivity, and introduces an expert-merging method for efficient deployment.

Findings

ModernALBERT achieves state-of-the-art performance on GLUE, SQuAD-v2, and BEIR.

MoL restores expressivity lost in parameter sharing, matching larger models.

Expert-merging compresses MoL into a single adapter with minimal accuracy loss.

Abstract

Parameter sharing in recursive transformers reduces model size but collapses layer-wise expressivity. We propose Mixture of LoRAs (MoL), a lightweight conditional-computation mechanism that inserts Low-Rank Adaptation (LoRA) experts inside a shared feed-forward network (FFN). MoL enables token-conditional weight-space modulation of the shared FFN without untying backbone parameters, unlike prior approaches that add fixed or externally attached adapters. We pretrain a modernised recursive architecture, ModernALBERT, integrating rotary embeddings, GeGLU, FlashAttention, and a distillation-based initialisation. Across GLUE, SQuAD-v2, and BEIR, ModernALBERT (50M--120M) achieves state-of-the-art performance among compact models and surpasses larger fully parameterised baselines. We also propose an expert-merging procedure that compresses MoL into a single adapter at inference while preserving accuracy, enabling efficient deployment. Our results show that conditional weight-space modulation effectively restores the expressivity lost under aggressive parameter sharing in recursive transformers.