ICL-KB: Charting the Failure Frontier of Few-Shot LLMs

In-context learning (ICL) with large language models (LLMs) has unlocked powerful few-shot capabilities, but it also exposes sharp knowledge boundaries where models fail in systematic ways. We introduce a layered framework, the In-Context Learning Knowledge Boundary (ICL-KB), which characterizes these failures through three concentric sources. Specifically, parametric overriders that dominate contextual cues, fragility in how models interpret and rely on prompt composition, and inference collapse when reasoning deteriorates under weak or misleading context. Unlike prior surveys that catalog LLM capabilities or provide high-level overviews of ICL, our work organizes existing empirical and theoretical findings through three complementary analytical lenses. Internal probing reveals how context–model conflicts manifest in hidden representations. External stress tests, including controlled benchmarks and adversarial prompts, expose the brittleness of ICL behaviors. Boundary extension examines mitigation strategies that expand the region in which ICL remains stable and reliable. Grounded in examples from knowledge-graph question answering, anomaly detection, and concept drift, the survey links ICL failure modes to long-standing challenges in data mining. These insights establish ICL-KB as a coherent, mechanistic roadmap for advancing reliable, interpretable, and theoretically grounded few-shot learning with LLMs.