New research suggests that using expert personas in AI prompting may hurt accuracy. A study from USC examines the effectiveness of persona-based prompting in AI and introduces PRISM, a technique to harness the benefits of expert personas while avoiding the drawbacks.
Researchers have discovered that the popular technique of using persona-based prompting in artificial intelligence, where users instruct AI models to adopt the persona of an expert, may be counterproductive, particularly for tasks requiring factual accuracy. This method, which involves prompts like 'You are an expert machine learning programmer,' has been in use since 2023.
A recent study, 'Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM,' conducted by researchers from the University of Southern California (USC), reveals that the effectiveness of persona-based prompting is highly task-dependent. While it can enhance performance in alignment-dependent tasks such as writing, role-playing, and safety, it tends to degrade results in pretraining-dependent areas like mathematics and coding. The underlying reason appears to be that simply instructing a model to behave like an expert does not impart actual expertise or add any new knowledge to its existing training data. In fact, such instructions may hinder the model's ability to access and utilize the facts it has already learned during pretraining. This research highlights a nuanced understanding of how these models function and how best to interact with them for optimal results across different applications. \The study employed the Measuring Massive Multitask Language Understanding (MMLU) benchmark to evaluate the impact of persona-based prompting. The results showed that when the AI model was tasked with multiple-choice questions, the expert persona consistently underperformed the base model across all subject categories, with an overall accuracy of 68.0 percent compared to the base model's 71.6 percent. This suggests that persona prefixes can activate the model's instruction-following mode at the expense of its ability to recall factual information. However, persona-based guidance did prove beneficial in steering the model toward responses that align with specific requirements, such as those related to safety. For example, a 'Safety Monitor' persona significantly improved the model's refusal rates in safety benchmarks. The researchers emphasized that instructing an AI to assume the role of an expert programmer will likely not improve the quality or utility of the generated code. They suggest that more granular and specific requirements related to project architecture, user interface preferences, or tool selection might be more helpful in guiding the model toward generating code that meets the user's needs. This indicates that while the concept of persona-based prompting is not inherently flawed, the way it is implemented and the types of personas employed are crucial to its success.\To address the limitations of expert personas, the researchers proposed a technique called PRISM (Persona Routing via Intent-based Self-Modeling). This method aims to leverage the benefits of expert personas without the drawbacks. PRISM involves a mechanism that retains the base model for generating outputs reliant on pretrained knowledge. A LoRA adapter is activated when persona-based behaviors are expected to improve the outcome; otherwise, the model reverts to its unmodified state. The decision on when to use the persona is made by a 'gate' learned by the model. The design of PRISM avoids the trade-offs associated with other approaches, such as prompt-based routing, which applies expert personas at inference time, and supervised fine-tuning, which incorporates behavior directly into the model's weights. One of the researchers, Hu, noted that while the research provided specific insights into the use of expert personas, it doesn't give a universal approach. He suggests, based on the findings, that specificity in prompts is more important when alignment and rules are the goals, and that for tasks that require factuality, it's often best to avoid adding any directives
AI Prompting Expert Personas LLM PRISM
United States Latest News, United States Headlines
Similar News:You can also read news stories similar to this one that we have collected from other news sources.
Best Collagen Creams for Mature Skin: Expert-Recommended PicksDiscover the top dermatologist-recommended collagen creams to help reduce fine lines, improve hydration, and boost skin elasticity. This article reviews several options, including affordable picks, night creams, and K-beauty favorites, with detailed pros and cons for each product.
Read more »
Mass cattle cull in Russia feared to be from bioweapons research plant 'leak'Thousands of animals have been destroyed amid furious protests from farmers in Siberia with authorities giving differing accounts as to the reason why.
Read more »
Gardening Expert Urges Planting Five Seeds in April for a Successful HarvestGardening expert Benedict Vanheems, known as Grow Veg, recommends planting five key seed types in April to maximize harvest success. He emphasizes the importance of utilizing the optimal sowing window as warmer weather arrives and shares specific advice on planting carrots, courgettes, and cucumbers, including tips on variety selection, pest control, and optimal growing conditions.
Read more »
OMV Backs Austrian SAF Research Hub to Accelerate Aviation DecarbonizationOMV has partnered with the Technical University of Leoben to establish a unique European research facility aimed at accelerating the development and commercialization of sustainable aviation fuels.
Read more »
Research reveals potential new strategy for targeting deadly small cell neuroendocrine cancersUCLA researchers have uncovered a hidden weakness in some of the deadliest cancers, revealing a potential new strategy for targeting tumors that have long resisted treatment.
Read more »
Labour Unveils Locations for Seven New Towns, Promising 1.5 Million New HomesThe Labour Party announces the locations for seven new towns as part of its ambitious plan to build 1.5 million homes. These developments, which will include affordable housing and extensive community amenities, are intended to revitalize several areas across England and address the nation's housing crisis.
Read more »