Researchers have proposed a new framework called PSL to improve large language models (LLMs) in predictive political question answering. They argue that existing methods treat external resources as knowledge-based evidence and neglect prediction-relevant signals. The authors identify two key signals: actor stances that capture issue-specific preferences, and high-order structure signals that capture indirect dependencies among political actors. PSL converts semi-structured political records into inference-oriented evidence for LLMs by extracting stance signals from question-relevant actor records and learning structure-aware actor representations. --- Why it matters: This matters to researchers in AI because it addresses a limitation of existing methods for augmenting LLMs with external data, potentially improving the accuracy of predictive political question answering. By incorporating both stance and structure signals, PSL could enhance the ability of LLMs to reason about complex political relationships and make more informed predictions. Source: https://arxiv.org/abs/2608.21218