ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning

TL;DR

ReMix introduces a reinforcement learning-based routing mechanism for Mixture-of-LoRAs in LLM finetuning, addressing imbalance issues and significantly improving performance over existing methods.

Contribution

The paper proposes a novel non-learnable routing design with an unbiased reinforcement learning gradient estimator for Mixture-of-LoRAs, enhancing their expressive power.

Findings

ReMix outperforms state-of-the-art methods in parameter-efficient finetuning.

The non-learnable routing ensures balanced utilization of LoRAs.

The reinforcement learning approach scales well with increased training compute.

Abstract

Low-rank adapters (LoRAs) are a parameter-efficient finetuning technique that injects trainable low-rank matrices into pretrained models to adapt them to new tasks. Mixture-of-LoRAs models expand neural networks efficiently by routing each layer input to a small subset of specialized LoRAs of the layer. Existing Mixture-of-LoRAs routers assign a learned routing weight to each LoRA to enable end-to-end training of the router. Despite their empirical promise, we observe that the routing weights are typically extremely imbalanced across LoRAs in practice, where only one or two LoRAs often dominate the routing weights. This essentially limits the number of effective LoRAs and thus severely hinders the expressive power of existing Mixture-of-LoRAs models. In this work, we attribute this weakness to the nature of learnable routing weights and rethink the fundamental design of the router. To address this critical issue, we propose a new router designed that we call Reinforcement Routing for Mixture-of-LoRAs (ReMix). Our key idea is using non-learnable routing weights to ensure all active LoRAs to be equally effective, with no LoRA dominating the routing weights. However, our routers cannot be trained directly via gradient descent due to our non-learnable routing weights. Hence, we further propose an unbiased gradient estimator for the router by employing the reinforce leave-one-out (RLOO) technique, where we regard the supervision loss as the reward and the router as the policy in reinforcement learning. Our gradient estimator also enables to scale up training compute to boost the predictive performance of our ReMix. Extensive experiments demonstrate that our proposed ReMix significantly outperform state-of-the-art parameter-efficient finetuning methods under a comparable number of activated parameters.