Google AI introduces AGREE – a machine learning framework that allows LLMs to substantiate claims in their answers and provide accurate citations

Maintaining the accuracy of large language models (LLMs), such as GPT, is crucial, particularly in cases that require factual accuracy, such as news reporting or educational content creation. Despite their impressive capabilities, LLMs tend to generate plausible but non-objective information, known as “hallucinations,” typically when faced with open-ended queries that require extensive knowledge of the world. Google ai researchers introduced AGREE to address the problem of “hallucinations,” where LLMs generate a response that is objectively incorrect, meaningless, or disconnected from the input message.

Existing approaches to preventing hallucinations in LLMs mainly include two methods: post hoc citation and cue-based grounding. Post hoc citations involve adding citations after generating responses, often using natural language inference (NLI) models. However, this method relies heavily on the knowledge included in the LLM and faces challenges with facts beyond its training data. While the prompt-based foundation takes advantage of the instruction-following and in-context learning capabilities of LLMs, it is often ineffective, particularly in real-world scenarios that require high objective accuracy.

The proposed solution, AGREE (Adaptation for GRounding EnhancEment), introduces a learning-based framework that allows LLMs to substantiate their answers themselves and provide accurate citations. AGREE takes a holistic approach by combining learning-based adaptation and test-time adaptation (TTA). During training, AGREE refines LLMs using synthetic data from unlabeled queries, allowing them to substantiate their claims by adding citations to their answers. AGREE uses an iterative inference strategy during testing time, allowing LLMs to actively search for more information based on self-generated citations, helping them improve their answers again and again.

In the training stage, AGREE involves collecting synthetic data from unlabeled queries, retrieving relevant passages from trusted sources using a retrieval model, and fine-tuning a base LLM to substantiate its claims. The adjustment process uses an NLI model to judge the support of each claim and add citations accordingly. Experiments on five data sets demonstrate the effectiveness of AGREE in improving database and citation accuracy compared to baseline methods. AGREE outperforms post hoc and cue-based dating approaches, achieving relative improvements of more than 30% in grounding quality. Additionally, AGREE can work with out-of-domain data, suggesting its robustness across different question types, including out-of-domain knowledge. The inclusion of TTA in AGREE also leads to improvements in both the substantiation and correctness of responses.

In conclusion, AGREE has effectively improved the issue of hallucinations in LLMs by working on its feasibility and verifiability. By allowing LLMs to substantiate their answers and provide accurate citations, AGREE improves their reliability, particularly in domains that require high factual accuracy. AGREE's approach of combining learning-based adaptation with test-time adaptation provides a robust solution that performs better than current approaches and can be used on a wide range of data sets. Overall, AGREE has the potential to promote reliable language models suitable for real-world applications that require high factual accuracy.

Review the Paper and Blog. All credit for this research goes to the researchers of this project. Also, don't forget to follow us on twitter.com/Marktechpost”>twitter. Join our Telegram channel, Discord channeland LinkedIn Grabove.

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

Don't forget to join our 43k+ ML SubReddit | Also, check out our ai Event Platform