StanceBank: An Efficient Reasoning Framework for Large-Scale Stance Detection on Social Media

Abstract

Stance detection underpins a wide range of large-scale data analytics applications on social media, including opinion monitoring, policy evaluation, and misinformation tracking. While prompting-based methods leveraging Large Language Models (LLMs) have achieved strong zero-shot performance, they invariably perform per-sample reasoning, re-deriving target-specific knowledge for every input. This pattern increases request-path API calls and becomes a key bottleneck when applied to the millions of opinionated posts generated daily. We observe that real-world stance corpora exhibit pronounced target clustering—a small number of targets account for the vast majority of samples—and that the target-level reasoning required within each cluster is largely sample-invariant. Motivated by this structural property, we propose StanceBank, a two-stage framework that explicitly decouples target-level reasoning from sample-level reasoning. In the offline Bank Construction stage, an LLM is prompted with a small set of annotated samples per target to elicit a structured, reusable reasoning framework capturing stance expression patterns, related entities, common pitfalls, and recommended reasoning steps. In the online Inference stage, each sample retrieves the corresponding framework and predicts its stance with a single LLM call; for unseen targets, the system falls back to Plan-and-Solve reasoning and can incrementally extend the Bank. Experiments on three benchmark datasets (SemEval-2016, P-Stance, VAST) and two LLM backbones (GPT-4o-mini and Claude Haiku 4.5) demonstrate that StanceBank matches strong reasoning baselines such as Plan-and-Solve in accuracy while reducing average request-path calls by 17% to 34%. When an additional plan-to-framework conversion is charged to every miss, the fixed-target benchmarks retain a total-call advantage, whereas open-target VAST does not. Our results identify target recurrence as the key condition for cost-efficient reasoning reuse.