TAP-LLM: An Executable Attribution Framework for Temporal Event Graph Prediction
Abstract
Although temporal event graph predictors can infer future relational events from historical sequences, their scores provide limited evidence about which historical events support a particular output. We propose TAP-LLM, an executable attribution framework for temporal event graph prediction. Rather than treating explanations as free-form text or static reasoning paths, TAP-LLM represents each attribution as a program containing candidate event identifiers (IDs) and predefined operations. The program specifies evidence selection, history perturbation, score comparison, and greedy evidence compression, thereby connecting large language model (LLM) planning to predictor-side verification. We instantiate the main protocol with RE-GCN on ICEWS14s and ICEWS05-15 and add preliminary cross-predictor verification with RE-Net. Budget-matched retrieval and temporal-occlusion baselines, target-level uncertainty analysis, and a cost audit distinguish program planning from retrieval and verification. TAP-LLM produces approximately two-event evidence chains with about 99% program validity; our claims concern predictor-grounded fidelity and compactness rather than human-perceived interpretability.