Telling an AI model 'You are a skilled programmer' can actually decrease its programming abilities.
There's a persistent rumor that telling an AI, 'You are an expert in this field,' improves its performance. A study investigating this rumor has been published on the preprint server arXiv.
[2603.18507] Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM
Telling an AI model that it's an expert makes it worse • The Register
Previous research has shown that AI performance improves when it is instructed with the information that 'you are an expert on the task you are performing.' For example, an AI instructed to pretend to be a bird expert was better at describing birds than an AI instructed to pretend to be a car expert.
This kind of information is widespread, and there are even prompt guides that suggest starting by making the AI aware that it is an expert.
Zhi Zhao Hu and colleagues at the University of Southern California investigated this rumor.
Hu et al. provided six different AI models, including Llama-3.1-8B and Qwen2.5-7B, with multiple different prompts and verified their performance in several benchmark tests.
The prompts were divided into simple and complex types. For example, if the simple prompt was 'You are a software engineer,' the complex prompt might be 'You are a data extraction and information retrieval specialist with deep expertise in natural language processing and document analysis. You have extensive experience in extracting accurate and relevant information from tables, web pages, and complex documents.' The complex prompts were designed to make the AI aware of its expertise.
As a result, in the 'MT-Bench' benchmark test, which evaluates performance through multiple interactions, improvements in output quality with complex prompts were observed in areas such as writing and reasoning. On the other hand, quality was found to have actually decreased in areas such as coding, mathematics, and humanities.
In the 'Measuring Massive Multitask Language Understanding (MMLU)' test, which assesses a wide range of knowledge, overall accuracy decreased, and it was found that telling people 'you are an expert' actually led to worse results.
This is thought to be because telling an AI to be an expert in a particular field does not actually grant it that expertise. Hu et al. stated, 'Instructing an AI to be an expert may cause it to use its ability to follow instructions instead of recalling facts, which it would normally use.'
This experiment revealed that prompts encouraging the AI to act like an expert degrade performance in coding and mathematics, but conversely, 'AI alignment,' which adjusts the AI to operate in accordance with human ethics, is suggested to actually improve performance. In particular, significant improvements were seen in the 'JailbreakBench' benchmark test, which verifies whether the AI can extract unethical content.
Mr. Hu stated, 'Based on the results of this study, telling an AI 'you are a skilled programmer' does not improve the quality or usefulness of its code. Rather, we believe that providing detailed project requirements is more helpful in generating code that meets the user's needs.'
Related Posts:
in AI, Posted by log1p_kr