AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and Reflective Prompting for Abductive Event Reasoning

2026-03-04

AI Summary

This work addresses systematic errors in multi-label causal reasoning arising from incomplete causal chains, recency bias, and salience-induced distortions. To mitigate these issues, the authors propose a three-stage abductive reasoning framework that integrates graph-structured retrieval, large language model (LLM)-based abductive generation, and a novel reflective prompting mechanism, further enhanced by post-hoc consistency constraints to improve inference accuracy. The study innovatively designs an evolving reflective prompting strategy that uncovers and validates three shared inductive biases across diverse LLM families, demonstrating that such errors are systemic rather than model-specific. Evaluated on SemEval-2026 Task 12, the approach achieves state-of-the-art performance with a 0.95 accuracy. Cross-model analysis across 14 architectures quantifies a 51% reduction in spurious causal attributions, providing strong empirical evidence for the prevalence of these systematic biases.

Technology Category

• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Machine Learning: Causal Learning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

We present a winning three-stage system for SemEval 2026 Task~12: Abductive Event Reasoning that combines graph-based retrieval, LLM-driven abductive reasoning with prompt design optimized through reflective prompt evolution, and post-hoc consistency enforcement; our system ranks first on the evaluation-phase leaderboard with an accuracy score of 0.95. Cross-model error analysis across 14 models (7~families) reveals three shared inductive biases: causal chain incompleteness, proximate cause preference, and salience bias, whose cross-family convergence (51\% cause-count reduction) indicates systematic rather than model-specific failure modes in multi-label causal reasoning.

Problem

• abductive reasoning
• causal reasoning
• inductive biases
• multi-label reasoning
• event reasoning

Innovation

• graph-based retrieval
• abductive reasoning
• reflective prompting
• consistency enforcement
• inductive bias