Aiyappa et al., "Zero-shot stance detection in practice" (2026)

Rachith Aiyappa, Shruthi Senthilmani, Jisun An, Haewoon Kwak, and Yong-Yeol Ahn, PeerJ Computer Science 12, e3540 (2026)
DOI | arXiv | Code

@article{aiyappa2026stance,
    author = {Rachith Aiyappa and Shruthi Senthilmani and Jisun An and Haewoon Kwak and Yong-Yeol Ahn},
    title = {Zero-shot stance detection in practice: Insights on training, prompting, and decoding with a capable lightweight {LLM}},
    journal = {PeerJ Computer Science},
    volume = {12},
    pages = {e3540},
    doi = {10.7717/peerj-cs.3540},
    year = {2026},
}

We examine how FlanT5-XXL performs at identifying stances in tweets without task-specific training. The zero-shot approach can match or outperform state-of-the-art benchmarks, including fine-tuned models. We reveal how instruction sensitivity, decoding methods, prompt complexity, and linguistic elements like negations affect performance, and discover a positivity bias that may explain performance variations across different decoding approaches.