The field of natural language creation has been completely transformed by large-scale language models (LLMs). The traditional fine-tuning approach for responding to downstream tasks requires access to the LLM’s parameters, which limits its use with powerful black-box LLMs that only provide an API (such as ChatGPT). For this reason, recent research has focused on prompting techniques that dictate generated results by providing many task-specific instructions and demonstrations, where prompts can have a significant impact on outcome outcomes. Therefore, it has been demonstrated that careful design is required.
Prompts are in principle a flexible method, but the methods commonly used today are rather rigid. But that is not the case with language learning. People can improve their language skills by receiving and responding to positive and negative feedback.
A new study by Northeastern University of China, Microsoft Research Asia, Microsoft Azure Translation, and NiuTrans Research revisits output deficiencies to determine whether and how LLMs improve their deliberative abilities. , is being asked to learn how to discover. To help identify errors before generation, they designed a new prompt template called Deliberate then Generate (DTG) that contains instructions and possible output.
Candidate determination is an important part of DTG design. Using data from a second baseline system is an easy option as its output is usually of high quality and requires only minor adjustments to be used effectively. Therefore, effective deliberation cannot be promoted. Researchers suggest using text unrelated to the source material, such as random text selections or null strings. DTG successfully taps into LLM’s deliberative powers without relying on other text generation systems to provide examples of modifications, making it easy to adapt to a wide variety of text production jobs with only minor changes to the prompts. Adaptable. From a psychological point of view, the work is inspired by the canonical case of language acquisition, considering the negative evidence in the development of language ability.
The team conducted extensive experiments and showed that the proposed DTG prompts reliably improved model performance compared to traditional prompts in GPT3.5 (text-DaVinci-003) and GPT4. This is true for 7 text generation tasks and over 20 datasets. Machine translation, simplification, and common-sense authoring are just some of the text generation tasks for which DTG-prompted GPT achieves state-of-the-art performance on large datasets. The proposed DTG prompt allows intentional ability and error avoidance prior to generation, as shown by extensive ablation studies and statistical error analysis.
The researchers plan to leverage task-specific domain knowledge in future studies to further improve the effectiveness of DTG prompts.
please check out paper. don’t forget to join 23,000+ ML SubReddit, Discord channeland email newsletterShare the latest AI research news, cool AI projects, and more. If you have any questions regarding the article above or missed something, feel free to email us. Asif@marktechpost.com
🚀 Check out 100’s of AI Tools at the AI Tools Club
Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her bachelor’s degree at the Indian Institute of Technology (IIT), Bhubaneswar. She is a data her science enthusiast and has a keen interest in the range of applications of artificial intelligence in various fields. She is passionate about exploring new advances in technology and its practical applications.
