Divide, Reweight, and Conquer: A Logit Arithmetic Approach for In-Context Learning
This work addresses efficiency and scalability issues in ICL for LLM users, though it is incremental as it builds on existing ICL methods with novel optimizations.
The paper tackles performance degradation and high computational costs in In-Context Learning (ICL) for Large Language Models by proposing LARA and B-LARA, which use logit-based ensembling and parallelization to improve accuracy and memory efficiency, as demonstrated by outperforming baselines on BBH and MMLU benchmarks.
In-Context Learning (ICL) emerges as a key feature for Large Language Models (LLMs), allowing them to adapt to new tasks by leveraging task-specific examples without updating model parameters. However, ICL faces challenges with increasing numbers of examples due to performance degradation and quadratic computational costs. In this paper, we propose Logit Arithmetic Reweighting Approach (LARA), a novel framework that enhances ICL by using logit-based ensembling of multiple demonstrations. Our approach divides long input demonstrations into parallelizable shorter inputs to significantly reduce memory requirements, and then effectively aggregate the information by reweighting logits of each group via a non-gradient optimization approach. We further introduce Binary LARA (B-LARA), a variant that constrains weights to binary values to simplify the search space and reduces memory usage by filtering out less informative demonstration groups. Experiments on BBH and MMLU demonstrate that LARA and B-LARA outperform all baseline methods in both accuracy and memory efficiency. We also conduct extensive analysis to show that LARA generalizes well to scenarios of varying numbers of examples from limited to many-shot demonstrations.