Zhongqiao Lin

Bowen Zheng, Ming Ma, Zhongqiao Lin et al.

Large language models are computationally expensive due to their deep structures. Prior research has shown that intermediate layers contain sufficient information to generate accurate answers, leading to the development of early-exit algorithms that reduce inference costs by terminating computation at earlier layers. However, these methods often suffer from poor performance due to misalignment between intermediate and output layer representations that lead to decoding inaccuracy. To address these challenges, we propose SPADE (SPace Alignment DEcoding), a novel decoding method that aligns intermediate layer representations with the output layer by propagating a minimally reduced sequence consisting of only the start token and the answer token. We further optimize the early-exit decision-making process by training a linear approximation of SPADE that computes entropy-based confidence metrics. Putting them together, we create a hybrid early-exit algorithm that monitors confidence levels and stops inference at intermediate layers while using SPADE to generate high-quality outputs. This approach significantly reduces inference costs without compromising accuracy, offering a scalable and efficient solution for deploying large language models in real-world applications.

Bowen Zheng, Ming Ma, Zhongqiao Lin et al.

Large Language Models (LLMs) have demonstrated remarkable abilities, one of the most important being in-context learning (ICL). With ICL, LLMs can derive the underlying rule from a few demonstrations and provide answers that comply with the rule. Previous work hypothesized that the network creates a task vector in specific positions during ICL. The task vector can be computed by averaging across the dataset. It conveys the overall task information and can thus be considered global. Patching the global task vector allows LLMs to achieve zero-shot performance with dummy inputs comparable to few-shot learning. However, we find that such a global task vector does not exist in all tasks, especially in tasks that rely on rules that can only be inferred from multiple demonstrations, such as categorization tasks. Instead, the information provided by each demonstration is first transmitted to its answer position and forms a local task vector associated with the demonstration. In some tasks but not in categorization tasks, all demonstrations' local task vectors converge in later layers, forming the global task vector. We further show that local task vectors encode a high-level abstraction of rules extracted from the demonstrations. Our study provides novel insights into the mechanism underlying ICL in LLMs, demonstrating how ICL may be achieved through an information aggregation mechanism.