IMPACT OF MODEL SIZE AND PROMPTING STRATEGY ON ZERO- AND FEW-SHOT PERFORMANCE IN OPEN-SOURCE LANGUAGE MODELS

What determines the capabilities of open-source language models: their parameter count or the manner in which they are prompted?To comprehensively distinguish these effects,we evaluate a diverse range of instruction-tuned models, including Flan-T5 checkpoints (small, base, large), and recent architectures with extended context windows,across a substantial scaled evaluation that encompasses hundreds of articles and diverse NLP tasks. Each model is subjected to multiple prompting regimes (zero-shot and few-shot with varying numbers of examplars), while controlled input lengths and prompt phrasings are maintained. Automatic scoring(ROUGE1/2/L,accuracy,macroF1)is complemented by multi-rater human evaluations that assess factuality, coherence, and faithfulness. The results demonstrate a pronounced interaction: scaling parameters consistently enhances baseline (zero-shot) performance, but the advantage of in-context demonstrations is significantly influenced by the alignment between prompt length,input size,and available context window. On short-context tasks such as Named Entity Recognition (NLI),well-selected exemplars substantially improve accuracy for larger models.Conversely,on long-context tasks like summarization, adding demonstrations can negatively impact performance by displacing critical input tokens a finding corroborated across multiple architectures and datasets. We propose a refined capacity context alignment principle: exemplars are only beneficial if the models context window and parameter scale can simultaneously accommodate them without compromising source information.These findings challenge conventional prompt engineering practices and provide practical, statistically supported recommendations for optimizing LLM deployment under real-world budget and resource limitations.