Beyond the Code: Google's New System for LLM Reliability, Anthropic's Breakthrough, Xi Jinping Chatbot

Welcome to the 33rd edition of LLMs: Beyond the Code!

In this edition, we'll explore:

• 谷歌 is developing frameworks to ensure accurate source citation in LLM outputs, enhancing reliability.
• Anthropic improves control over LLM outputs by breaking down the AI black box.
• A new LLM from Chinese researchers designed to mimic the political philosophies of Xi Jinping.

Let's jump right into the latest developments in generative AI!

Google's AGREE Addresses Factual Errors in LLMs

Google's AGREE framework significantly enhances the reliability of large language models (LLMs) by improving their ability to accurately cite sources in their responses.

• AGREE trains LLMs to support their responses with precise citations, enhancing accuracy and reliability.
• The framework generates initial responses from a base LLM, then assesses them using a natural language inference (NLI) model to ensure factual grounding.
• Responses confirmed by the NLI model to be factually supported are used to fine-tune the LLM, teaching it to autonomously include accurate citations in future outputs.

This approach marks a significant step forward in making LLMs more reliable by ensuring their responses are grounded in verifiable facts.

For more information, read here.

Meet XiBot, China's Political Chatbot

Researchers in China have developed XiBot, a new AI chatbot based on President Xi Jinping's philosophies, currently under internal testing.

• XiBot uses a LLM fine-tuned on government documents, including the book "Xi Jinping Thought on Socialism with Chinese Characteristics for a New Era," to facilitate interactions that reflect Xi’s political views.
• By integrating these philosophies into a conversational AI, it provides a novel means for users to explore and understand complex political ideas through interactive dialogue.
• The primary use of XiBot is to make the principles of Xi Jinping’s governance approach more accessible to the public.

By facilitating direct interaction with these ideas, the chatbot helps users gain a deeper understanding of Chinese political thought and policies in an engaging manner.

For more information, read here.

Anthropic Unlocks the Black Box of LLMs

Anthropic researchers have made great progress in understanding the "black box" nature of large language models through their model, Claude.

• The team used a method called "dictionary learning" to study Claude's outputs. They mapped specific neuron combinations to different ideas or "features."
• By adjusting the activation levels of these neuron combinations, researchers could control the model's output.
• For example, they lowered the activations for combinations linked to negative outputs (like unsafe code or violent content) to reduce these occurrences. They increased activations for combinations tied to positive features (like safety protocols or accurate information) to make them more prominent.
• This method allows researchers to understand and improve the model's internal workings, making it safer and more functional.

This breakthrough helps unravel the "black box" nature by making it clearer how specific neuron activations lead to certain outputs.

For more information, read here.

AI Legal Tools Still Make Costly Mistakes Despite RAG

Even with RAG, AI tools from LexisNexis and Thomson Reuters are transforming law practice, but their reliability remains a concern.

• Researchers from Stanford RegLab and HAI found that AI legal research tools still make mistakes more than 17% of the time.
• The study showed that these tools can give incorrect legal information and wrong citations.
• Even with RAG, these systems have problems like difficulty finding the right legal documents, retrieving irrelevant authorities, and agreeing with wrong user assumptions.

This study highlights that hallucinations in large language models remain a significant concern, pointing to the urgent need for improved testing and transparency for their use in legal settings.

For more information, read here.

Thanks for tuning in to this week's edition of LLMs: Beyond the Code!

If you enjoyed this edition, please leave a like and feel free to share with your network.

See you next week!