Stance prompts stop being rewritten whole: six modules tuned one at a time, errors routed by label, four of five backbones gain

A stance prompt is no longer rewritten as one block of text; it is split into six functional modules, task definition, guidelines, favor definition, against definition, neutral definition, and output format, and updated one at a time in a fixed order. Earlier automatic prompt optimization was largely task-general and not organized around stance-label class structure, so guidelines, label definitions, and output constraints played undifferentiated roles during optimization. On SemEval-2016, modular optimization reaches higher mean Macro-F1 than full-prompt optimization for four of five backbones, with the largest gap on Mistral at 3.49 points.

Misclassification feedback is routed by gold label: the favor definition only sees errors whose gold label is favor, the against definition only sees against errors, and the guidelines module aggregates errors from all three classes. When errors from all three classes are pooled and sampled equally, the signal reaching each label module is diluted by errors from other classes. On SemEval-2016, class-conditioned routing gives higher mean Macro-F1 for four of five backbones, by 3.40 for Phi, 1.05 for GPT-OSS, 4.99 for Mistral, and 2.21 for Gemma, while Llama drops 0.15.

Optimization changes only the prompt text and never updates model parameters, so inference uses the selected prompt directly for classification. Fine-tuning requires updating weights, whereas prompt optimization keeps the change at the text level. Experiments cover four datasets, SemEval-2016, GunStance, COVID-19-Stance, and the new VaccineStance, across five open-weight models: Phi-4-14B, GPT-OSS-20B, Mistral-Small-3.2-24B, Gemma-3-27B, and Llama-3.3-70B.

The method is evaluated on four stance-detection datasets and five open-weight models, optimization updates no model weights, and the selected prompt needs no further optimization calls at inference. It targets social-media analysis settings that need stance classification, such as misinformation detection, political opinion analysis, and public health monitoring. VaccineStance is built from vaccine-related posts collected between 2013 and 2022, with 500 training, 500 validation, and 1,000 test instances labeled independently by three annotators.

Optimization trajectories show most models improving in early iterations, with smaller and more fluctuating gains later, and sharp drops followed by recovery on VaccineStance. The relative effectiveness of the two reasoning-based strategies, I+R and I+R+E, varies across models and datasets, and no configuration is better in every setting. Frozen-module ablations show module contributions vary substantially across backbones, and freezing the output format even yields small improvements for Gemma and Llama. The external reasoner GPT-5 improves 17 of 20 dataset-model combinations but drops Mistral by 5.82 points on SemEval. The authors also note that generalizability to other NLP tasks is unexplored, that optimized prompts are model- and domain-dependent, that the fixed module structure targets three labels and extending it to larger label spaces remains future work, and that API-based model updates may affect reproducibility.