Retrieval-augmented generation is a language model that answers by first looking up passages and then reading them. When one of those passages is partly false, current systems make an all-or-nothing choice. They either keep the passage, and the model repeats the false clause, or they throw the whole passage away, losing the true facts around it. Misinformation is usually exactly this mixed kind. A fake news story typically keeps the real event, people and place and distorts one number, one attribution, or one causal link. We propose to repair retrieved passages before the model reads them, using the oldest idea from logical belief revision: make the smallest change that removes the contradiction. Concretely, we split each passage into short atomic claims and find the smallest groups of claims that cannot all be true. Within each group we replace the least trusted claim with the strongest weaker version that is still consistent with the evidence, for example "the deal was signed in 2019" becoming "the deal was signed (date disputed)." We then retract any claims that were supported only by the revised one.
The bet is that a knowledge base that deletes whole entries loses a great deal of correct information, and so does a retrieval system that deletes whole passages. Database and ontology research measured that loss and found ways to avoid it. Retrieval systems have not taken up those methods. The proposal is falsifiable in a sharp way. If deleting only the contradicted atomic claims does as well as weakening them, then the belief-revision machinery adds nothing. We build the benchmark and the ablation so that this question gets a clear answer.
Core Claim
Given retrieved passages that mix true and false content, a repair step that keeps as much as possible while staying consistent will produce better answers than three alternatives: keeping the passages as they are, filtering whole passages, or picking between conflicting sources. It should also beat plain claim-level deletion on two kinds of questions. The first kind is answered by true material sitting next to a false span. The second kind depends on a false fact through a chain of inference. We will measure this as a frontier between two quantities. One is how many true reference claims the repaired context still supports, which we call retention. The other is how many false claims it still supports, which we call leakage. The claim is that weakening plus retraction of dependent claims moves the frontier outward. Lower leakage alone does not count.
Where This Sits
The motivating evidence comes from four places. In knowledge bases, deleting whole axioms or tuples destroys valid knowledge. Repairing Ontologies via Axiom Weakening rewrites a conflicting rule into a more specific one instead of dropping it, and on the taxrank ontology it keeps 0.82 of the inferable content versus 0.56 for deletion. Computing Approximate Query Answers over Inconsistent Knowledge Bases updates single attribute values instead of discarding whole records. A Semantic Approach for Iterated Revision in Possibilistic Logic names the "drowning effect," where resolving one high-priority conflict wipes out unrelated lower-priority beliefs. We reuse the operators these papers define (minimal inconsistent sets, weakening, value update, source priority). We apply them to claims extracted from text rather than to logical formulas, and we give up their formal guarantees in exchange.
On the misinformation side, MisSpans shows that true and false spans coexist inside single sentences. It also finds that 15 large language models score below 0.5 F1 at locating and typing the false spans. MisSpans is a detection benchmark only, and we turn its annotations into a repair-and-answer task. Its two span types map naturally onto our two operations. An "Inconsistency" span is a contradiction to revise. A "No mention" span is an unsupported addition to hedge rather than delete.
Three retrieval systems already handle conflicts, and each resolves them at a coarser unit than we propose. Astute RAG groups internal and retrieved knowledge into clusters by source and picks the more reliable cluster. CARE-RAG writes a free-text conflict report that the generator conditions on. TruthfulRAG converts passages into subject–relation–object triples and uses a spike in output entropy to switch which reasoning path the model follows. None of the three edits the evidence in a way that keeps the true remainder of a corrupted passage. None tracks which claims depended on a revised one. TruthfulRAG is the closest, but it selects whole paths and inherits the errors of its triple extraction.
Two diagnostic papers constrain the design. ECON finds that contradiction detectors lose up to 18.2% F1 moving from sentence to paragraph inputs, and that models resolve conflicts by chance in 24–38% of cases. This is why we detect conflicts between short claims, not between passages, and why source priority is decided explicitly rather than left to the generator. Belief Revision: The Adaptability of Large Language Models Reasoning shows frontier models scoring only 50–52% on its revision score, with chain-of-thought adding about 1%. This argues against expecting the generator to revise its beliefs inside its own context, and is why we move revision into a separate step that runs before generation.
Research Program
The repair step has four stages. First, an extractor splits each retrieved passage into atomic claims, each aligned to its character span. Second, a claim-level natural language inference model (a classifier that labels a pair of statements as entailing, contradicting, or neutral) is run over claim pairs across passages, and optionally against claims elicited from the model's own memory. This builds a conflict graph whose minimal inconsistent sets are the groups to resolve. Third, each claim gets a priority, analogous to epistemic entrenchment, computed from how many independent sources corroborate it and from source metadata. In each minimal inconsistent set, the lowest-priority claim is revised.
Revision proposes candidate weakenings in order from strongest to weakest. These are: replace the value with a range or the disputed set of values, generalize the entity to its type, keep the claim as attributed ("source X reports…"), and finally drop it. The strongest candidate that no longer contradicts the higher-priority claims is kept. Fourth, retraction removes or hedges any claim that is entailed by the original false claim but by nothing else in the repaired context. This is the step that addresses the drowning effect from the opposite direction: we remove consequences of the false claim and leave unrelated claims untouched. The repaired claims are written back into the passage in place, so their original wording and position are kept, with disputed values marked. Formally, among repairs $R$ that leave the context consistent, we pick
$$R^* = \arg\max_{R\ \text{consistent}} \sum_{c \in C} w_c \, \mathrm{info}(c, R),$$
where $C$ is the set of extracted claims, $w_c$ is claim $c$'s priority, and $\mathrm{info}(c,R) \in \{0, \text{partial}, 1\}$ scores whether $c$ survives intact, weakened, or not at all. A greedy approximation is adequate because minimal inconsistent sets in news passages are small.
We also build the evaluation, which we call MisSpans-QA. For each fake/real article pair in MisSpans, which is built from FakeNewsAMT, the fake article is placed among retrieved passages under three conditions. In the first, the real article or a partial corroborator is present. In the second, only weak corroboration is present. In the third, there is no corroboration, and the only correct behavior is to hedge. For each item we write three kinds of questions. Collateral questions are answered by true spans next to a false one. Target questions ask directly about the false span, and the correct response is the corrected value or an abstention. Dependent questions are answered only through an inference that passes through the false span. ECON's factoid perturbations (entity swaps and negations at graded intensity) give a second, controlled source of the same three question types.
Baselines
The naive reference points are a closed-book model with no retrieval and vanilla retrieval-augmented generation that feeds the passages unchanged. The coarse-deletion baseline is a passage-level filter: a DeBERTa-V3-large cross-encoder fine-tuned for inference as in From RAG to Reality, which drops any passage labeled contradictory. That paper reports weighted F1 0.58 on SciHal. This is the strongest simple filter in the file and is precisely the "delete the whole entry" policy whose cost we claim to measure.
The strongest existing conflict-handling systems are Astute RAG and CARE-RAG. CARE-RAG reports an Exact Match gain of 23.6% over standard retrieval-augmented generation on Natural Questions with LLaMA-3.2-8B. TruthfulRAG reports 85.0% versus 67.3% accuracy on RealtimeQA with GPT-4o-mini. We found no reported result on Lacuna for any of these systems on MisSpans, ECON-derived question answering, or FaithEval, so all of them are rerun by us.
The key ablations isolate our ingredients. Claim-level deletion removes the contradicted atomic claims and does nothing else. Weakening without retraction tests the dependency step on its own. Prompted rewriting, where a model is asked to "correct this passage," uses the same model as our pipeline.
Fairness is enforced by holding the following fixed across all methods: the same retriever, top-k, generator, and claim-level inference model. Two generators are used, an open 8B model and one commercial model, because Astute RAG is reported to degrade on small models. Every method gets the same budget of tuning examples. We report tokens and model calls per query alongside accuracy.
Beating the baselines means three things on MisSpans-QA. On collateral and dependent questions combined, accuracy must be higher than claim-level deletion and Astute RAG, with non-overlapping bootstrap confidence intervals. Target-question accuracy must be no worse than the best baseline. The full retention–leakage curve must dominate the curves of the other methods.
Why not just…?
Why not just drop the contradicted atomic claims? This is the real threat. If weakening and retraction add little, the honest conclusion is that claim-level filtering is all that is needed, and we will report that. The prediction that separates the two approaches is specific. Deletion should do as well on collateral questions, but worse on dependent questions, where false consequences survive, and worse on target questions, where it gives no corrected or disputed value to report. The no-corroboration condition should show the largest gap, because deletion has nothing to delete there unless it removes the claim outright. The weakening-without-retraction ablation then shows which of the two ingredients carries the effect.
Why not just ask a strong model to rewrite the passage? The evidence that models do this reliably is poor. MisSpans finds that 15 models, including frontier ones, locate false spans at under 0.5 F1. ECON finds chance-level resolution in 24–38% of conflicts. The Belief-R results show prompting strategies add about 1% to revision ability. Because the comparison is run with the same model and the same budget, any advantage we show comes from structure rather than from model strength.
Why not just filter whole passages with a good inference classifier? This approach throws away exactly the material that collateral questions need. ECON also shows that passage-level detection loses up to 18.2% F1 compared with sentence-level detection. In addition, the From RAG to Reality classifier confuses unverifiable claims with entailed ones in 32.58% of its errors, so unsupported "No mention" additions pass through. The retention–leakage curve measures this tradeoff directly.
Why not just use source-aware conflict prompting like Astute RAG, or a bigger model? FaithEval shows that larger models revert to their memorized priors when the context is counterfactual: GPT-4o falls from about 96% to 47.5%. Cluster selection also has to keep or reject a mixed passage as a whole. If Astute RAG matches us on collateral questions with the commercial model, that will show the advantage of our method is limited to smaller models, and we will scope the claim accordingly.
Proposed Experiments and Decisive Checks
The main experiment runs every method on MisSpans-QA and on ECON-derived question answering, broken down by question type and corroboration condition. We report answer accuracy plus retention and leakage of reference claims, each judged by a held-out claim-level inference model and audited by human annotation on a sample of 300 items.
Two transfer checks follow. The first applies the method to FaithEval's Inconsistent and Counterfactual subsets, to test that repair does not cause the model to ignore legitimate counterintuitive context. The second applies it to RealtimeQA against TruthfulRAG's reported numbers. As a byproduct, the spans our pipeline revises are scored as MisSpans span predictions, compared against the reported sub-0.5 F1.
The decisive checks are two. First, the dependent-question gap between our method and claim-level deletion must be positive and must grow with inference depth. Second, our retention–leakage curve must lie strictly above deletion's at matched leakage. If both fail, the proposal reduces to claim-level filtering. If both hold, we will have shown that minimal-change repair from belief revision measurably reduces information loss in retrieval systems.
References
- Repairing Ontologies via Axiom Weakening
- Computing Approximate Query Answers over Inconsistent Knowledge Bases
- A Semantic Approach for Iterated Revision in Possibilistic Logic
- MisSpans: Fine-Grained False Span Identification in Cross-Domain Fake News
- Astute RAG: Overcoming Imperfect Retrieval Augmentation and Knowledge Conflicts for Large Language Models
- Rethinking All Evidence: Enhancing Trustworthy Retrieval-Augmented Generation via Conflict-Driven Summarization
- TruthfulRAG: Resolving Factual-level Conflicts in Retrieval-Augmented Generation with Knowledge Graphs
- ECON: On the Detection and Resolution of Evidence Conflicts
- Belief Revision: The Adaptability of Large Language Models Reasoning
- From RAG to Reality: Coarse-Grained Hallucination Detection via NLI Fine-Tuning