Google AI introduces AGREE: a machine learning framework that helps LLMs self-justify answer claims and provide accurate citations

Maintaining the accuracy of large language models (LLMs) such as GPT is critical, especially when factual accuracy is required, such as in news reporting or creating educational content. Despite their impressive capabilities, LLMs tend to generate plausible but non-factual information (called “hallucinations”) when faced with open-ended queries that require extensive world knowledge. Google AI researchers introduced AGREE to address the problem of “hallucinations,” where LLMs generate responses that are counterfactual, nonsensical, or disconnected from the input prompt.

Existing approaches to prevent hallucinations in LLMs mainly include two methods: post-quoting and prompt-based grounding. Post-quoting adds quotations after generating a response and often uses natural language inference (NLI) models. However, this method relies heavily on knowledge within the LLM's embeddings and faces challenges with facts outside of the training data. Prompt-based grounding leverages the prompt-following and in-context learning capabilities of LLMs, but is often ineffective, especially in real-world scenarios where high factual accuracy is required.

The proposed solution, AGREE (Adaptation for GRounding Enhancement), introduces a learning-based framework to enable LLMs to self-ground responses and provide accurate citations. AGREE employs a holistic approach that combines both learning-based adaptation and test-time adaptation (TTA). During training, AGREE uses synthetic data from unlabeled queries to fine-tune the LLM so that it can self-ground claims by adding citations to responses. At test time, AGREE uses an iterative inference strategy, which enables the LLM to proactively seek more information based on its self-generated citations, helping it improve answers iteratively.

In the training phase, AGREE collects synthetic data from unlabeled queries, uses a retriever model to retrieve relevant sentences from trusted sources, and fine-tunes the base LLM to self-ground its claims. The fine-tuning process uses the NLI model to determine the support for each claim and adds citations accordingly. Experiments across five datasets demonstrate that AGREE improves grounding and citation accuracy compared to baseline methods. AGREE outperforms prompt-based and post-citation approaches, with a relative improvement in grounding quality of over 30%. Furthermore, AGREE can also handle out-of-domain data, suggesting that it is robust across a range of question types that include out-of-domain knowledge. The inclusion of TTA in AGREE improves both grounding and answer accuracy.

In conclusion, AGREE effectively ameliorated the hallucination problem in LLMs by addressing the factuality and verifiability of LLMs. By enabling LLMs to self-ground responses and provide accurate citations, AGREE increases the reliability of LLMs, especially in domains where high factual accuracy is required. AGREE's approach, combining learning-based adaptation and test-time adaptation, provides a powerful solution that outperforms current approaches and can be used on a wide range of datasets. Overall, AGREE has the potential to facilitate reliable language models suitable for real-world applications where high factual accuracy is required.

Please check Papers and blogs. All credit for this research goes to the researchers of this project. Also, don't forget to follow us. twitter. participate Telegram Channel, Discord Channeland LinkedIn GroupsUp.

If you like our work, you will love our Newsletter..

Please join us 43,000+ ML subreddits | In addition, our AI Event Platform