Researchers from MBZUAI developed "uncertainty quantification heads" (UQ heads) to detect hallucinations in language models by probing internal states and estimating the credibility of generated text. UQ heads leverage attention maps and logits to identify potential hallucinations without altering the model's generation process or relying on external knowledge. The team found that UQ heads achieved state-of-the-art performance in claim-level hallucination detection across different domains and languages. Why it matters: This approach offers a more efficient and accurate method for identifying hallucinations, improving the reliability and trustworthiness of language models in various applications.
MBZUAI researchers developed FarSight, a plugin to reduce hallucinations in Multimodal Large Language Models (MLLMs). FarSight addresses the issue where MLLMs generate inaccurate text by losing focus on relevant image details, leading to snowball hallucinations. Testing on models like LLaVA-1.5-7B showed FarSight's effectiveness in reducing initial mistakes, thereby minimizing overall hallucinations. Why it matters: Improving the reliability of MLLMs is crucial for applications requiring high accuracy, enhancing their utility in various real-world scenarios.
MBZUAI researchers presented a new uncertainty quantification method at ACL to identify hallucinations in LLMs, called claim conditioned probability (CCP). CCP leverages the internal token probabilities generated by the LLM itself to highlight claims with low confidence. Unlike external fact-checking methods, CCP is computationally efficient as it uses probabilities already computed by the model. Why it matters: This research offers a practical approach to mitigate the impact of LLM hallucinations by highlighting potentially unreliable information, improving the trustworthiness of these models, especially for Arabic LLMs.
The paper introduces AraHalluEval, a new framework for evaluating hallucinations in Arabic and multilingual large language models (LLMs). The framework uses 12 fine-grained hallucination indicators across generative question answering and summarization tasks, evaluating 12 LLMs including Arabic-specific, multilingual, and reasoning-based models. Results show factual hallucinations are more common than faithfulness errors, with the Arabic model Allam showing lower hallucination rates. Why it matters: This work addresses a critical gap in Arabic NLP by providing a comprehensive tool for assessing and mitigating hallucination in LLMs, which is essential for reliable AI applications in the Arabic-speaking world.
MBZUAI Professor Preslav Nakov is researching methods to identify and combat the harmful uses of large language models in generating disinformation. He notes that disinformation, unlike fake news, is weaponized with the intent to persuade, not just to lie. His research focuses on the linguistic differences between human-written and machine-generated disinformation, such as the use of rhetorical devices in human propaganda. Why it matters: As AI-generated content becomes more prevalent, understanding and mitigating its potential for spreading disinformation is critical for maintaining trust and integrity in information ecosystems, especially during major election cycles.
This paper introduces DetectLLM-LRR and DetectLLM-NPR, two novel zero-shot methods for detecting machine-generated text using log rank information. Experiments across three datasets and seven language models demonstrate improvements of up to 3.9 AUROC points over state-of-the-art methods. The code and data for both methods are available on Github.
MBZUAI's Maxim Panov is developing uncertainty quantification methods to improve the reliability of language models. His work focuses on providing insights into the confidence level of machine learning models' predictions, especially in scenarios where accuracy is critical, such as medicine. Panov is working on post-processing techniques that can be applied to already-trained models. Why it matters: This research aims to address the issue of "hallucinations" in language models, enhancing their trustworthiness and applicability in sensitive domains within the region and globally.