In the deployment of Large Language Models (LLMs), “spurious correctness”—where answers are correct but reasoning contains errors—poses a critical risk by creating an illusion of reliability. While prior work on LLM confidence estimation focuses on answer-level or entire reasoning path confidence, these coarse-grained approaches fail to identify which specific parts of the reasoning contain errors. We propose a fine-grained confidence estimation framework that computes confidence scores for individual evidence triplets within reasoning chains, enabling precise localization of errors. Using carefully designed prompts, we generate answers, evidence in triplet format, and their respective confidence scores simultaneously, allowing automatic detection of spurious correctness patterns where partial evidence contains factual errors. Evaluated on both Japanese and English multi-hop QA benchmarks across multiple models from three model families representing different architectures and training approaches, we show that our approach exhibits superior calibration performance for evidence confidence and demonstrates effective ability to detect spurious correct answers (up to 0.84 on our primary discrimination metric). The consistent improvements across languages demonstrate the generalizability of our method. As a secondary benefit, joint generation of confidence scores improves answer confidence calibration by up to 43%. This prompt-based approach requires no model retraining and is immediately applicable to existing LLMs.

Recent LLMs show an impressive accuracy on one of the hallmark tasks of language understanding, namely Question Answering (QA). However, it is not clear if the correct answers provided by LLMs are actually grounded on the correct knowledge related to the question. In this paper, we use multi-hop QA datasets to evaluate the accuracy of the knowledge LLMs use to answer questions, and show that as much as 31% of the correct answers by the LLMs are in fact spurious, i.e., the knowledge LLMs used to ground the answer is wrong while the answer is correct. We present an analysis of these spurious correct answers by GPT-4 using three datasets in two languages, while suggesting future pathways to correct the grounding information using existing external knowledge bases.

We present JEMHopQA, a multi-hop QA dataset for the development of explainable QA systems. The dataset consists not only of question-answer pairs, but also of supporting evidence in the form of derivation triples, which contributes to making the QA task more realistic and difficult. It is created based on Japanese Wikipedia using both crowd-sourced human annotation as well as prompting a large language model (LLM), and contains a diverse set of question, answer and topic categories as compared with similar datasets released previously. We describe the details of how we built the dataset as well as the evaluation of the QA task presented by this dataset using GPT-4, and show that the dataset is sufficiently challenging for the state-of-the-art LLM while showing promise for combining such a model with existing knowledge resources to achieve better performance.

This paper presents a hybrid approach to the verification of statements about historical facts. The test data was collected from the world history examinations in a standardized achievement test for high school students. The data includes various kinds of false statements that were carefully written so as to deceive the students while they can be disproven on the basis of the teaching materials. Our system predicts the truth or falsehood of a statement based on text search, word cooccurrence statistics, factoid-style question answering, and temporal relation recognition. These features contribute to the judgement complementarily and achieved the state-of-the-art accuracy.