TLDR: T-REX (Table – Refute or Entail eXplainer) is a pioneering live, interactive tool that enables non-experts to verify textual claims against structured tabular data. Utilizing state-of-the-art Large Language Models, T-REX provides accurate, real-time, and transparent fact-checking across multimodal and multilingual tables. It addresses the limitations of existing solutions by offering an accessible interface, real-time reasoning with highlighted evidence, and ensuring user data privacy by not storing any information.
In today’s information-rich world, verifying claims against data is more crucial than ever. While much attention is given to text-based fact-checking, a significant challenge lies in validating textual claims against structured tabular data. This task, known as table fact-checking, requires a blend of linguistic understanding, logical reasoning, and numerical computation. Despite advancements driven by Large Language Models (LLMs), many existing solutions remain complex, often confined to academic settings, and inaccessible to those without technical expertise.
Enter T-REX, which stands for Table – Refute or Entail eXplainer. This innovative tool is designed to bridge the gap between advanced fact-checking technology and everyday users. T-REX is the first live, interactive system that allows anyone to verify claims against multimodal and multilingual tables using cutting-edge, instruction-tuned reasoning LLMs. Its core mission is to provide accurate and transparent fact-checking, empowering non-experts to combat misinformation effectively.
How T-REX Works: A Glimpse Behind the Scenes
T-REX operates on a modular framework, integrating several key components to deliver its powerful capabilities. Users can easily provide tables either by uploading CSV files or by inputting images, from which T-REX extracts data using advanced Optical Character Recognition (OCR) technology. Once a table is provided, along with a textual claim and an optional table title for context, the system’s inference engine takes over.
The development of T-REX involved extensive experimentation to find the most effective approach for claim verification. Researchers explored various strategies, including direct prompting of LLMs, generating code for verification, and using Retrieval-Augmented Generation to find relevant table content. After rigorous testing, the team found that a technique called ‘direct prompting with Chain-of-Thought (CoT)’ yielded the best results. This method not only achieved high accuracy (89% on the TabFact dataset) but also enhanced interpretability by guiding the LLM to output its reasoning steps, making the verification process transparent.
T-REX supports several state-of-the-art open-source LLMs, including Phi-4, Cogito v1 Preview, DeepSeek-R1-Distill-Qwen-7B, and Gemma 3, carefully selected to balance powerful reasoning abilities with practical hardware constraints.
Designed for the User: Accessibility and Transparency
The user interface of T-REX is built with accessibility and transparency at its forefront. Users can initiate the verification process by simply uploading or pasting a table and a claim. The system then streams its reasoning in real-time, generates a clear verdict (entailed or refuted), and visually highlights the relevant cells within the table preview. This highlighting feature is crucial for interpretability, allowing users to understand exactly which parts of the table the model used to reach its conclusion.
To further enhance usability, T-REX offers multilingual input and output in eight languages, making it globally accessible. It also provides exportable JSON outputs for advanced users, and automatically renders all table inputs as editable CSVs within the interface. This last feature is particularly useful for correcting any recognition errors that might occur with OCR inputs, ensuring data accuracy.
Under the Hood: Robust and Secure Implementation
T-REX is built on a robust FastAPI backend, orchestrating LLM inference, OCR processing, and data management. It leverages Ollama for both LLM inference and OCR tasks, and incorporates rate limiting to maintain stability under heavy use. Importantly, all inference and data handling are performed in memory, meaning no user data is stored or transmitted externally, ensuring privacy and security.
Also Read:
- Making AI Transparent: LLM-Powered Explanations for Knowledge Graph QA
- Boosting Information Extraction: A New AI Workflow Combines Language Models with Logic
The Future of Fact-Checking
T-REX represents a significant step forward in making advanced table fact-checking accessible to everyone. It provides accurate, real-time, and interpretable claim verification without compromising user data. Looking ahead, the developers plan to expand T-REX’s capabilities to include precise numerical computation, structured claim decomposition, and support for other table-centric tasks like table completion and semantic table retrieval.
To learn more about T-REX and explore its capabilities, you can read the full research paper here: T-REX Research Paper.
