Do different prompting methods yield a common task representation in language models?

2025-05-17

AI Summary

It remains unclear whether different prompting paradigms—few-shot demonstrations versus natural-language instructions—elicit consistent internal task representations in large language models. Method: Building on the function vector framework, we extend it for instruction-based prompts for the first time, integrating attribution-controlled experiments, layer-wise activation decomposition, and zero-shot/few-shot evaluation. Contribution/Results: We find that demonstrations and instructions activate largely non-overlapping model components, indicating divergent task representations. Instruction-specific function vectors are successfully extracted and improve zero-shot accuracy. Although the two prompting modalities rely on distinct computational pathways, they exhibit synergistic performance gains when combined. This work provides the first representation-level theoretical foundation for hybrid prompting (instruction + demonstration), advancing both the interpretability and principled design of prompting mechanisms.

Technology Category

• Natural Language Processing: Prompt Engineering / Prompting
• Machine Learning: Large Multimodal Models (LMMs)
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Demonstrations and instructions are two primary approaches for prompting language models to perform in-context learning (ICL) tasks. Do identical tasks elicited in different ways result in similar representations of the task? An improved understanding of task representation mechanisms would offer interpretability insights and may aid in steering models. We study this through function vectors, recently proposed as a mechanism to extract few-shot ICL task representations. We generalize function vectors to alternative task presentations, focusing on short textual instruction prompts, and successfully extract instruction function vectors that promote zero-shot task accuracy. We find evidence that demonstration- and instruction-based function vectors leverage different model components, and offer several controls to dissociate their contributions to task performance. Our results suggest that different task presentations do not induce a common task representation but elicit different, partly overlapping mechanisms. Our findings offer principled support to the practice of combining textual instructions and task demonstrations, imply challenges in universally monitoring task inference across presentation forms, and encourage further examinations of LLM task inference mechanisms.

Problem

• Compare task representations from different prompting methods
• Analyze model components used by instruction vs demonstration prompts
• Investigate overlap between mechanisms for varied task presentations

Innovation

• Generalize function vectors to instruction prompts
• Extract instruction function vectors for zero-shot accuracy
• Analyze demonstration- and instruction-based vector differences