Implicit Demonstration Augmentation for Robust and Stable In-Context Learning.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks without parameter updates by using contextual demonstrations.
• ICL performance is sensitive to demonstration quality, quantity, and order, leading to instability.
• Existing ICL methods lack robust mechanisms for demonstration optimization.

Purpose of the Study:

• To introduce demonstration augmentation as a novel approach to enhance ICL performance in LLMs.
• To develop methods that improve the stability and accuracy of LLM predictions through enriched demonstrations.
• To address the limitations of current ICL techniques concerning demonstration selection and integration.

Main Methods:

• Proposed Implicit Demonstration Augmentation-based ICL (IDAICL), which enriches demonstrations using deep feature distributions from the entire demonstration set.
• Developed Domain-aware IDAICL (D-IDAICL) using a hypernetwork to adaptively select domain-specific knowledge for augmenting demonstrations based on test sample representations.
• Theoretically demonstrated that IDAICL converges to a novel form of logit calibration with infinite augmented samples.

Main Results:

• Both IDAICL and D-IDAICL significantly improved overall and worst-case accuracy across multiple tasks and eight LLMs.
• The proposed methods enhanced the robustness and predictive capabilities of LLMs.
• Performance fluctuations due to demonstration variations (order, templates) were mitigated, and class imbalance was handled effectively.

Conclusions:

• Demonstration augmentation is a promising direction for improving ICL in LLMs.
• IDAICL and D-IDAICL offer effective solutions for unstable ICL performance and enhance LLM generalization.
• These methods contribute to more reliable and robust LLM applications through optimized in-context learning.