Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Misconfidence-based Demonstration Selection for LLM In-Context Learning

Shangqing Xu, Chao Zhang

TL;DR

Misconfidence-based Demonstration Selection for LLM In-Context Learning addresses the sensitivity of in-context learning to demonstration quality. The authors propose In-Context Reflection (ICR), which uses a misconfidence metric to identify and replace demonstrations that reveal gaps between the LLM's output distribution and the task mappings, without requiring extra supervision. Across 13 subtasks from five task sets, ICR yields about a 4% average improvement and demonstrates cross-task generalization, with ablations indicating robustness to several design choices. The approach offers a scalable, label-distribution-preserving strategy for crafting few-shot prompts that enhances ICL's robustness in diverse domains.

Abstract

In-context learning with large language models (LLMs) excels at adapting to various tasks rapidly. However, its success hinges on carefully selecting demonstrations, which remains an obstacle in practice. Current approaches to this problem either rely on hard-to-acquire external supervision or require frequent interactions with LLMs, resulting in high costs. We propose a new method called In-Context Reflection (ICR) to overcome these challenges. ICR strategically selects demonstrations to reduce the discrepancy between the LLM's outputs and the actual input-output mappings. Specifically, ICR starts with a random set of initial demonstrations, then iteratively refines it. In each step, it analyzes a pool of candidate examples and identifies the ones most likely to challenge the LLM's current understanding, measured by a new metric called misconfidence. These most confusing examples are then selected to replace the less informative demonstrations in the current set. Our comprehensive evaluation across five diverse datasets encompassing 13 subtasks shows the efficacy of ICR. Compared to existing methods, ICR achieves an average performance boost of 4%, while demonstrating remarkable cross-task generalization capabilities.

Misconfidence-based Demonstration Selection for LLM In-Context Learning

TL;DR

Misconfidence-based Demonstration Selection for LLM In-Context Learning addresses the sensitivity of in-context learning to demonstration quality. The authors propose In-Context Reflection (ICR), which uses a misconfidence metric to identify and replace demonstrations that reveal gaps between the LLM's output distribution and the task mappings, without requiring extra supervision. Across 13 subtasks from five task sets, ICR yields about a 4% average improvement and demonstrates cross-task generalization, with ablations indicating robustness to several design choices. The approach offers a scalable, label-distribution-preserving strategy for crafting few-shot prompts that enhances ICL's robustness in diverse domains.

Abstract

In-context learning with large language models (LLMs) excels at adapting to various tasks rapidly. However, its success hinges on carefully selecting demonstrations, which remains an obstacle in practice. Current approaches to this problem either rely on hard-to-acquire external supervision or require frequent interactions with LLMs, resulting in high costs. We propose a new method called In-Context Reflection (ICR) to overcome these challenges. ICR strategically selects demonstrations to reduce the discrepancy between the LLM's outputs and the actual input-output mappings. Specifically, ICR starts with a random set of initial demonstrations, then iteratively refines it. In each step, it analyzes a pool of candidate examples and identifies the ones most likely to challenge the LLM's current understanding, measured by a new metric called misconfidence. These most confusing examples are then selected to replace the less informative demonstrations in the current set. Our comprehensive evaluation across five diverse datasets encompassing 13 subtasks shows the efficacy of ICR. Compared to existing methods, ICR achieves an average performance boost of 4%, while demonstrating remarkable cross-task generalization capabilities.
Paper Structure (23 sections, 2 equations, 7 figures, 6 tables, 1 algorithm)

This paper contains 23 sections, 2 equations, 7 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Formulation
  4. Method
  5. In-Context Learning by Bridging Discrepancy
  6. In-Context Reflection (ICR)
  7. Experiment
  8. Settings
  9. Datasets
  10. Baselines
  11. Evaluation and Implementation Details
  12. Same-Task Evaluation
  13. Different-Task Evaluation
  14. Ablation
  15. Multiple ICR Iterations
  16. ...and 8 more sections

Figures (7)

  • Figure 1: A overview of our method. We aim to leverage the exact discrepancy between LLM's knowledge and task input-output mappings. Then we select demonstrations that best bridge such discrepancies.
  • Figure 2: An overview of In-Context Reflection (ICR). We first use an initialized prompt to get LLM prediction for each candidate, then calculate the misconfidence score $\psi$ to measure the discrepancy between LLM and task. After that, we rerank all candidates according to $\psi$, and replace part of the prompt with the top-ranked candidates, obtaining a refined prompt.
  • Figure 3: Different-task evaluation accuracy of ICR's prompt on GLUE's tasks. Each number shows the performance gain compared with same-task uniform sampling. On all tasks, ICR received comparable (sometimes even superior) results.
  • Figure 4: Performance gain of ICR on random prompt for 5 iterations. ICR produces a positive effect most times but is unstable.
  • Figure 5: Visualization of the result from demonstrations with different misconfidence levels on a) GLUE-MRPC b) Poem Sentiment. Demonstrations with larger misconfidence lead to better in-context performance.
  • ...and 2 more figures