Enhancing LLMs in Predictive Political QA with Semi-Structured Data introduces PSL, a framework designed to improve how Large Language Models (LLMs) predict political outcomes, such as voting behavior. The researchers, Yinan Liu, Zihan Zhou, Zichun Jin, Xinyu Wang, Bin Wang, and Xiaochun Yang, argue that standard LLMs struggle with predictive political tasks because external resources are often treated as simple factual knowledge rather than evidence for reasoning.
Addressing Predictive Reasoning
Predictive political questions—like how an actor will vote on a future bill—cannot be solved by factual lookup alone because the event has not yet occurred. The authors identify two signals missing from current LLM augmentation methods: actor stances, which reflect specific ideological preferences, and high-order structure signals, which capture indirect dependencies between political actors. By using semi-structured data, the PSL framework aims to convert raw records into inference-oriented evidence that helps LLMs reason through these complex political dynamics.
The Dual-View Framework
PSL processes political data through two distinct views to provide the LLM with better context:
Semantic View: The framework constructs "political behavior profiles" in JSON format for each actor. It uses an iterative "ladder retrieval" process to select the most relevant records based on the specific question. A smaller, distilled model then extracts the actor's stance from these records, providing the LLM with a concise summary of the actor's preferences.
Vector View: The framework builds an interaction graph where actors are linked by shared records (e.g., voting on the same bill). It propagates information across this graph to create "neighbor-aware" actor representations. These representations are then co-embedded with the question focus, allowing the LLM to account for group-level influence and indirect political relationships.
Experimental Performance
The researchers tested PSL across three datasets—RCVP, ICEWS, and StaId—using several LLMs, including Llama-3.1-8B-Instruct, Mistral-7B-Instruct, and Deepseek-7B-Chat. According to the results, PSL consistently outperformed baseline methods, including those based on knowledge graphs and profile-based simulations. Ablation studies confirmed that both the stance signals and the structure signals contributed to the framework's overall effectiveness, with the combination providing the strongest results.
Considerations for Implementation
The framework relies on the availability of semi-structured political records, such as legislative voting data and diplomatic interaction logs. While the authors note that this format preserves more context than knowledge graph triples, the system requires a multi-step pipeline—including ladder retrieval, knowledge distillation for the stance model, and graph-based representation learning—before the final injection into the LLM. The authors utilized LoRA (Low-Rank Adaptation) to fine-tune the LLMs for this specific reasoning task.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!