I've trained a Large Language Model (LLM) using the RAG method to answer user queries. However, I'm facing an issue where the model lacks confidence in its answers and changes them based on user input, even when the initial response is correct.

For example, when asked "What is the capital of France?", the model correctly responds with "Paris." However, if the user replies "No, it's Berlin," the model accepts this incorrect response and later provides "Berlin" as the capital of France when asked again.

I've tried using different prompt templates to reinforce answer consistency, but the issue persists. How can I improve the model’s robustness and prevent it from altering correct answers based on user responses? Any suggestions or guidance would be greatly appreciated.

I've trained a Large Language Model (LLM) using the RAG method to answer user queries. However, I'm facing an issue where the model lacks confidence in its answers and changes them based on user input, even when the initial response is correct.

For example, when asked "What is the capital of France?", the model correctly responds with "Paris." However, if the user replies "No, it's Berlin," the model accepts this incorrect response and later provides "Berlin" as the capital of France when asked again.

I've tried using different prompt templates to reinforce answer consistency, but the issue persists. How can I improve the model’s robustness and prevent it from altering correct answers based on user responses? Any suggestions or guidance would be greatly appreciated.

• nlp
• langchain
• large-language-model
• aiml

1 Answer 1

You can add a prompt that enforces the model to prioritize earlier answers to ensure consistency. For example, you may ask the model to validate if its new answer conflicts with its prior knowledge, and only change the answer if its new input is significantly more reliable.

A possible prompt template like “Are you confident if this new answer is correct based on your knowledge?”.

However, when generating responses, you can adjust the model’s temperature and sampling strategies. A higher temperature often leads to more varied outputs, while a lower temperature results in more deterministic answers. By controlling these parameters, you can increase the model's confidence.