A new preprint describes an AI method that aims to stop deceptive information from being carried through the linked steps of a multi-hop answer. In its controlled ideal setting, the method, called ReliableRAG, reported the highest Exact Match (EM) scores among the tested systems on all three benchmarks: 55.10% on HotPotQA, 49.50% on 2WikiMultiHopQA and 31.60% on MuSiQue.
The ideal setting used a fixed contrast. Each question received three diverse low-credibility documents supporting the same erroneous answer. High-credibility documents were assigned a fixed credibility score of 10 and low-credibility documents a score of 1. In a second, evaluator-generated setting, credibility scores could range from 0 to 10 instead of being fixed in advance.
With evaluator-generated scores, the reported EM values were 40.30% for HotPotQA, 33.80% for 2WikiMultiHopQA and 19.50% for MuSiQue. Those figures are lower than the ideal-setting results, but they remain scores from the same controlled benchmark exercise, not a measure of accuracy in a live service.
Breaking information into smaller checks
The system’s central move is to break retrieved material into structured triples, the paper’s term for individual information units. It evaluates each triple for relevance to the question and for credibility, retains reliable non-redundant triples, and then builds the reasoning chain autoregressively — adding one step after another.
The experiments used HotPotQA, 2WikiMultiHopQA and MuSiQue, three multi-hop question-answering benchmarks built around Wikipedia-sourced documents. Researchers randomly sampled 1,000 questions from each test set for evaluation and 100 questions from each development set for hyperparameter tuning.
The main answer-quality measures were Exact Match and F1 score. The paper also tested how performance changed as the number of injected low-credibility documents varied.
Where the method held up
Across that stress test, ReliableRAG was reported to outperform six compared methods while showing minimal performance drops as the number of low-credibility documents changed. The evaluator-generated results showed a similar pattern.
In an ablation analysis, removing Triple Extraction was associated with a reported 34.1% EM drop on HotPotQA, while excluding Triple Evaluation was associated with a reported 15.6% decrease. The paper does not clarify whether those percentages are relative changes or percentage-point differences.
The reported balance between the two scoring priorities — semantic relevance and credibility — came at α* = 0.4. A separate Indirect Effect analysis reported positive effects in 70.5% of cases.
What the scores do not settle
The strongest numbers come with a narrow frame. The evidence is limited to computational tests on three benchmark datasets, using Wikipedia-sourced documents and controlled, LLM-generated low-credibility content. The study does not show how the method would perform with naturally occurring or adaptive misinformation, continuously changing sources, or a user-facing deployment.
Reliability judgments also depend on the models doing the extracting and evaluating. In the reported evaluator analysis, AUC, a discrimination measure, was 0.86 for document-level assessments and 0.67 for triple-level assessments with GLM-4-Flash across the three test sets. The paper reports no confidence intervals, p-values or repeated-run variability, so the size and stability of the differences are not formally quantified.
Depth was another constraint in the sensitivity analysis. When the maximum reasoning-chain length exceeded 4, EM was reported as virtually stagnant or slightly declining. The evaluation covered selected generators, encoders, evaluators and selectors, leaving broader transfer uncertain.
The manuscript says ReliableRAG’s source code is available online anonymously, although that access was not independently verified. The front matter identifies the work as arXiv:2608.25487v1 [cs.CL], dated 26 August 2026.
Taken together, the paper offers benchmark evidence for checking relevance and credibility at the level of small information units before constructing a reasoning chain. It does not establish calibrated credibility scores, reduced real-world misinformation harm, or reliable performance beyond the tested datasets, injection procedures and model configurations. Those questions remain open despite the strong ideal-setting scores.
Paper data and sources
Original title: ReliableRAG: Combating Misinformation in Retrieval-Augmented Generation via Reliability-Guided Reasoning Chains
Authors: Jinpu Jiang, Xuan Wu, Wenhao Song et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text