Hallucination Detection Using Evidence Retrieval and Natural Language Inference
This paper introduces TruthLens, an evidence-based pipeline that detects hallucinations in large language model responses by extracting individual claims, retrieving evidence via a hybrid BM25 and MiniLM approach, and verifying them with a DeBERTa-v3 NLI model, achieving moderate accuracy on the FEVER dataset while highlighting limitations in generalization and evidence coverage.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast landscape of modern artificial intelligence, large language models have become remarkably skilled at writing. They can compose fluent essays, answer complex questions, and tell coherent stories with a confidence that often feels indistinguishable from human expertise. Yet, beneath this polished surface lies a persistent flaw: these systems sometimes generate statements that sound plausible but are factually incorrect, unsupported, or entirely made up. This phenomenon, known as hallucination, poses a significant challenge because a single answer might contain a mix of true facts and false ones. If a system labels an entire response as simply "wrong" or "right," it misses the nuance of which specific parts are trustworthy. To solve this, researchers are turning to a method of verification that treats every sentence as a separate claim, checking it against a library of known facts to see if the evidence supports, contradicts, or simply fails to address the statement.
A team of researchers at the Indian Institute of Technology Roorkee has built a prototype system called TruthLens to tackle this problem. Rather than guessing whether an answer is true, TruthLens acts like a meticulous fact-checker that breaks a long answer down into individual sentences. For each sentence, the system searches a massive digital library of text to find relevant passages. It uses two different strategies to find these passages: one looks for exact word matches, while the other understands the meaning behind the words, even if the phrasing is different. The system then combines the results of these two searches to pick the most likely candidate evidence. Finally, it uses a specialized language model to compare the original claim against the chosen evidence, deciding if the evidence proves the claim, disproves it, or leaves the question unanswered.
The researchers tested this approach using a dataset of nearly twenty thousand examples derived from a standard fact-checking benchmark. They first measured how well their system could find the right evidence. When using only the word-matching search, the system found the correct evidence in about half of the cases. When using the meaning-based search, it improved slightly. However, when they combined both methods, the system successfully located the relevant evidence in 56.51% of the cases at its first guess, and in over 65% of cases if it looked at the top five or ten suggestions. This hybrid approach proved to be the most effective way to gather potential proof, showing that relying on just one search method is not enough to capture all the necessary information.
Once the system found the best evidence, it had to decide what that evidence meant. The researchers tested several configurations to see which setup produced the most accurate verdicts. They found that feeding the single best piece of evidence to the verification model worked better than trying to aggregate multiple pieces of evidence. On a separate set of two thousand examples that the system had never seen before, this final setup correctly classified the claims 62.25% of the time. While this is a solid performance for a prototype, the results also highlighted where the system struggles. The model was particularly good at identifying when a claim was supported or contradicted, but it found it difficult to distinguish between a claim that was simply unsupported by the available text and one that was actually false. In these uncertain cases, the system often hesitated, labeling them as "unknown" rather than making a definitive call.
The study also revealed that the system's success is tightly bound to the quality and scope of the evidence it can access. The researchers noted that their prototype relies on a specific, limited collection of text, and it cannot verify claims about events or facts that do not exist within that library. Furthermore, the system sometimes fails when a claim requires connecting multiple pieces of information or resolving references like pronouns, which can confuse the sentence-by-sentence analysis. In an attempt to test if their method could work on a different type of dataset designed for general question-answering, the researchers found that the system did not transfer well, suggesting that the tools built for one specific type of fact-checking task may not work universally without significant adjustment.
Ultimately, TruthLens demonstrates that it is possible to build a pipeline that breaks down AI answers and checks them against a database of facts, but it also exposes the limits of current technology. The system is not a magic wand that can instantly reveal the absolute truth of any statement; rather, it is a tool that shows what can be verified given the available evidence. It successfully identifies when a claim is supported or contradicted by the text it can find, but it also honestly admits when the evidence is missing or insufficient. The researchers conclude that while their hybrid retrieval and verification method works, the biggest hurdles remain in handling complex reasoning, resolving unclear references, and expanding the system's ability to work across different types of data. For now, the system serves as a functional prototype that offers a glimpse into how machines might one day help humans navigate the flood of information, provided we remember that the final judgment still requires human oversight.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.