EDGELM: Edge Demonstrations for Language Models’ Table Understanding

Large language models (LLMs) are increasingly used for table-centric data wrangling tasks, such as entity resolution, schema alignment, and anomaly detection. While in-context learning is a powerful tool for these tasks, its performance is highly sensitive to the quality of the provided demonstrations. Standard retrieval methods often select examples based on similarity to the query, which can reinforce a model's existing biases or likely predictions rather than helping it navigate difficult decision boundaries. This paper investigates whether selecting demonstrations that specifically highlight these boundaries can improve performance without requiring model retraining or task-specific engineering.

The authors propose EdgeLM, a task-agnostic framework for demonstration selection. Instead of relying solely on similarity, EdgeLM retrieves two types of "edge evidence":

• Data Edges: These are nearby examples in the instance space that possess different ground-truth labels than the query. By including these, the model is exposed to local label contrasts that define the decision boundary.
• Model Edges: These are nearby examples that the deployed LLM previously misclassified during a zero-shot evaluation. These examples serve as corrective evidence, forcing the model to confront its own specific failure modes.

EdgeLM combines these two types of evidence into a single prompt, keeping the total number of demonstrations within a fixed budget. The method is evaluated across five data wrangling tasks, fifteen datasets, and five different LLMs (both open-weight and proprietary).

EdgeLM consistently outperforms or matches standard retrieval methods across all tested models and tasks. The authors demonstrate that the two forms of edge evidence—data edges and model edges—provide complementary benefits. While similarity-based retrieval is effective for establishing general context, the addition of boundary-focused demonstrations helps the model resolve ambiguous cases where simple similarity would otherwise lead to incorrect predictions. The framework is shown to be robust across different model architectures and scales, requiring only a small, one-time precomputation step to identify model-specific errors.