Bin Yu

Soufiane Hayou, Nikhil Ghosh, Bin Yu

In this paper, we show that Low Rank Adaptation (LoRA) as originally introduced in Hu et al. (2021) leads to suboptimal finetuning of models with large width (embedding dimension). This is due to the fact that adapter matrices A and B in LoRA are updated with the same learning rate. Using scaling arguments for large width networks, we demonstrate that using the same learning rate for A and B does not allow efficient feature learning. We then show that this suboptimality of LoRA can be corrected simply by setting different learning rates for the LoRA adapter matrices A and B with a well-chosen ratio. We call this proposed algorithm LoRA$+$. In our extensive experiments, LoRA$+$ improves performance (1-2 $\%$ improvements) and finetuning speed (up to $\sim$ 2X SpeedUp), at the same computational cost as LoRA.

Soufiane Hayou, Nikhil Ghosh, Bin Yu

Low-Rank Adaptation (LoRA) is a widely used finetuning method for large models. Its small memory footprint allows practitioners to adapt large models to specific tasks at a fraction of the cost of full finetuning. Different modifications have been proposed to enhance its efficiency by, for example, setting the learning rate, the rank, and the initialization. Another improvement axis is adapter placement strategy: when using LoRA, practitioners usually pick module types to adapt with LoRA, such as Query and Key modules. Few works have studied the problem of adapter placement, with nonconclusive results: original LoRA paper suggested placing adapters in attention modules, while other works suggested placing them in the MLP modules. Through an intuitive theoretical analysis, we introduce PLoP (Precise LoRA Placement), a lightweight method that allows automatic identification of module types where LoRA adapters should be placed, given a pretrained model and a finetuning task. We demonstrate that PLoP consistently outperforms, and in the worst case competes, with commonly used placement strategies through comprehensive experiments on supervised finetuning and reinforcement learning for reasoning.

Ruixuan Miao, Xu Lu, Cong Tian et al.

The commonly used Reinforcement Learning (RL) model, MDPs (Markov Decision Processes), has a basic premise that rewards depend on the current state and action only. However, many real-world tasks are non-Markovian, which has long-term memory and dependency. The reward sparseness problem is further amplified in non-Markovian scenarios. Hence learning a non-Markovian task (NMT) is inherently more difficult than learning a Markovian one. In this paper, we propose a novel \textbf{Par}allel and \textbf{Mod}ular RL framework, ParMod, specifically for learning NMTs specified by temporal logic. With the aid of formal techniques, the NMT is modulaized into a series of sub-tasks based on the automaton structure (equivalent to its temporal logic counterpart). On this basis, sub-tasks will be trained by a group of agents in a parallel fashion, with one agent handling one sub-task. Besides parallel training, the core of ParMod lies in: a flexible classification method for modularizing the NMT, and an effective reward shaping method for improving the sample efficiency. A comprehensive evaluation is conducted on several challenging benchmark problems with respect to various metrics. The experimental results show that ParMod achieves superior performance over other relevant studies. Our work thus provides a good synergy among RL, NMT and temporal logic.