If you ask a single person for a piece of trivia, you might receive a highly confident but completely fabricated answer. However, if you ask ten independent people and eight of them provide the exact same specific detail, your confidence in that detail naturally increases. This simple human heuristic forms the backbone of a powerful paradigm in data quality, information extraction, and generative AI: redundancy-based verification.
As Wang-Chiew Tan notes in the Letter from the Special Issue Editors (IEEE Data Eng. Bull. 2011), truth discovery in multi-source environments relies on evaluating the relative weight of competing claims by measuring the frequency of agreement across independent sources. This corroboration mechanism acts as a quantitative proxy for trust, allowing systems to reconcile contradictory information without any prior knowledge of source reliability.
The Statistical Asymmetry of Truth and Noise
Why does frequency correlate so strongly with truth? The answer lies in a fundamental statistical asymmetry: truth is consistent, while errors are idiosyncratic.
When an information extraction system misinterprets a sentence or a large language model (LLM) hallucinates a fact, the resulting error is typically a random artifact of that specific context or sample. These errors rarely repeat in the exact same way across independent trials. In contrast, factual truth is anchored to reality, meaning independent sources or generation paths will repeatedly converge on it.
Doug Downey et al. formalized this behavior in their URNS model (A Probabilistic Model of Redundancy in Information Extraction), demonstrating that the frequency of repeated observations can be used to mathematically estimate the probability of a fact being correct. Because correct facts repeat far more often than random extraction errors, high redundancy serves as an excellent proxy for confidence.
This phenomenon is clearly visible when scaling up knowledge base construction. Dawn Lawrie et al. observed in HLTCOE Participation in TAC KBP 2016: Cold Start and EDL that the frequency of extracted values for an entity attribute typically follows a power law distribution. While extraction noise results in a long tail of unique, incorrect values that appear only once, the valid fact stands out because it receives the highest number of document attestations.
Cross-Source and Cross-System Triangulation
To maximize the power of redundancy, systems must ensure that the repeating information comes from genuinely independent paths. True validation happens when we triangulate across different data views or system architectures.
Early web-scale extraction systems exploited the structural redundancy of the internet. For instance, Mark Craven et al. demonstrated in Learning to Extract Symbolic Knowledge from the World Wide Web how hypertext environments offer multiple independent paths—such as personal homepages, project directories, and course syllabi—to verify the same underlying fact. By requiring agreement across these distinct sources, systems can filter out localized errors.
This multi-view approach was extended by Andrew Carlson et al. with the Never-Ending Language Learner (NELL) (Toward an Architecture for Never-Ending Language Learning), which enforced consensus between entirely different data views, such as unstructured free text and structured web tables.
Redundancy can also be measured across entirely different software architectures. Vidhoon Viswanathan et al. (Stacked Ensembles of Information Extractors for Knowledge-Base Population) built meta-classifiers that evaluate cross-system agreement. They discovered that when multiple independent extraction architectures identify the exact same fact from identical text spans (matching down to the specific source document and character offsets), the probability of that claim being correct increases substantially.
Consensus-Driven Ranking in Question Answering
In open-domain Question Answering (QA), the final challenge is often selecting the single correct answer from a noisy pool of retrieved candidates. Traditional systems rank candidates based purely on the retrieval score of the source passage. However, a more robust signal is the degree of agreement within the candidate pool itself.
Nan Duan demonstrated in Minimum Bayes Risk based Answer Re-ranking for Question Answering that the degree of agreement between candidate answers provides a powerful signal for correctness, as similar candidates reinforce each other's probability. This consensus-based approach allows a system to identify the correct answer even when the initial top-ranked candidate is wrong.
This concept of evidence reinforcement was further developed by Jeongwoo Ko et al. in A Probabilistic Framework for Answer Selection in Question Answering. Their framework increases the confidence score of a candidate answer by summing its similarity scores against all other extracted candidates in the set, treating the presence of redundant or synonymous candidates as a validation signal.
Similarly, Bernhard Kratzwald et al. (RankQA: Neural Question Answering with Answer Re-Ranking) and Jinxi Xu et al. (TREC 2002 QA at BBN: Answer Selection and Confidence Estimation) utilized the frequency of identical answer strings across diverse retrieved snippets to filter out high-scoring but isolated false positives.
The Modern Frontier: Fact-Level Voting in LLMs
With the rise of Large Language Models, the challenge has shifted from extracting static text snippets to mitigating fluid model hallucinations. Modern systems have adapted the principle of informational redundancy from selecting answers to synthesizing them through fine-grained, fact-level voting.
Instead of relying on a single model generation, systems sample a pool of independent responses. In retrieval-augmented generation, Chao-Chun Hsu et al. (Answer Generation for Retrieval-based Question Answering Systems) showed that generative models can naturally leverage informational redundancy across multiple retrieved candidates to identify correct answers, treating consistent repetition as a signal of factual grounding.
For long-form generation, evaluating entire answers as single units is often too coarse, as a response might mix accurate facts with subtle hallucinations. To solve this, Ante Wang et al. (Improving LLM Generations via Fine-Grained Self-Endorsement) introduced a self-endorsement framework. By sampling multiple responses and letting the model vote on individual, fine-grained facts rather than entire answers, the system can synthesize a final response that is substantially more accurate than any single original sample.
Taking this a step further, Raghuveer Thirukovalluru et al. developed Atomic Self-Consistency (ASC) (Atomic Self-Consistency for Better Long Form Generations). Instead of picking the single best candidate, ASC breaks multiple generated responses down into their constituent "atomic facts," clusters them by semantic meaning, and merges the most consistent pieces into a single, comprehensive answer.
[Prompt] ───► Generate N Responses ───► Break into Atomic Facts ───► Cluster & Count Frequency ───► Merge High-Frequency Facts
Because factual details recur frequently across independent generations while idiosyncratic hallucinations vary wildly, this consensus-based merging strategy filters out errors without requiring internal access to the model's token probabilities.