Introduction

Every day, the world is shaped by an endless stream of events—economic fluctuations, policy decisions, natural disasters, technological breakthroughs. Yet behind every headline lies a deeper story: Why did this happen?

Understanding the causes behind events is crucial not only for humans making sense of the world, but also for intelligent systems tasked with interpreting it.

While large language models (LLMs) have demonstrated impressive capabilities in event extraction, summarization, and future prediction, they still fall short in abductive reasoning—inferring the most likely cause of a given outcome from incomplete or distributed information. This missing piece limits their application in high-stakes scenarios such as misinformation detection, policy impact assessment, and crisis response.

To address this challenge, we introduce Abductive Event Reasoning (AER), a novel shared task in SemEval 2026 that investigates LLMs’ ability to reason about real-world event causality. Given a specific event (e.g., “Cryptocurrency Market Prices Soar”) and a set of retrieved documents, the model must identify the most plausible and direct cause, such as “Government announces national cryptocurrency reserve”. This task mirrors how humans reason under uncertainty—by piecing together context, background knowledge, and likely hypotheses to form the best explanation.

Our team has built a high-quality dataset spanning politics, finance, and public emergencies, where causal options are carefully constructed and validated through both LLMs and human annotators. Through pilot experiments, we show that even state-of-the-art LLMs struggle with this task due to semantic distraction and a tendency to  plausible yet incorrect answers. These findings reveal an important limitation in current LLM reasoning capacity.

The AER task pushes the boundary of what it means for a language model to "understand". Beyond extracting or summarizing, it demands structured, context-grounded reasoning, rooted in both textual evidence and internal knowledge.  As a result, it holds broad relevance for AI transparency, explainability, and decision-making systems.

By participating in this task, researchers can contribute to advancing LLMs' reasoning capabilities, while also tackling challenges with far-reaching societal impact—from improving the interpretability of AI systems to helping journalists, analysts, and citizens trace back the causal chains behind today's most urgent events.