Enhancing LLM Faithfulness and Robustness via Prompting
To improve an LLM's robustness against inaccurate retrieved texts, one can design prompts that explicitly instruct the model to be more faithful to facts. Such prompts can also empower the LLM to abstain from answering a question if the provided information is deemed insufficient or incorrect, thus preventing the generation of unsupported claims.
0
1
Tags
Ch.3 Prompting - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Enhancing LLM Faithfulness and Robustness via Prompting
Improving a Faulty Retrieval-Augmented Chatbot
A team developing a Retrieval-Augmented Generation (RAG) system for a legal database finds that it occasionally generates incorrect legal interpretations. This happens when the system retrieves legal documents that are either irrelevant to the user's query or outdated. The team decides to implement a two-part solution. Which of the following options best exemplifies the dual-approach strategy for handling this inaccurate retrieval?
Improving Retrieval Accuracy in RAG
A team is working to reduce incorrect outputs from their Retrieval-Augmented Generation (RAG) system, which are caused by flawed retrieved documents. Match each of the two primary strategies they can employ with its corresponding description and goal.
Learn After
Example Prompt Instruction for Faithfulness and Abstention
Evaluating a Prompting Strategy for Factual Accuracy
A developer is building a question-answering system and provides a language model with the following instruction: 'You must base your answer strictly on the provided text. If the text does not contain enough information to answer the question accurately, you must respond with the exact phrase
Insufficient Information.' The model is then given the question 'What is the capital of Australia?' and the following text: 'Australia is a country and continent surrounded by the Indian and Pacific oceans. Its major cities are Sydney, Brisbane, Melbourne, and Perth.' Based on the developer's instruction, what is the most appropriate response from the model?Designing a Prompt for Factual Summarization