Imagine trying to verify whether a complex medical claim is true. If you hand this claim to a single, monolithic language model, it must simultaneously recall the exact scientific literature from its parameters and perform logical deduction. As Qing Zong et al. demonstrate in ComparisonQA, monolithic models frequently rely on semantic shortcuts or guess answers based on entity frequency rather than genuine factual recall. When confronted with rare or low-frequency facts, end-to-end architectures break down.
To solve this, the field of automated fact-checking has converged on a modular retrieval-verification pipeline. Instead of expecting a single model to execute search and reasoning simultaneously, the task is broken down into a sequence of specialized stages. This architecture isolates information retrieval from semantic inference, ensuring that final verdicts are explicitly grounded in verifiable source text.
The Anatomy of the Canonical Pipeline
The standard modular architecture decomposes open-domain fact-checking into discrete, sequential phases. As outlined by Zhijiang Guo et al. in A Survey on Automated Fact-Checking, the workflow typically follows three core stages:
[Input Claim] ──> [Document Retrieval] ──> [Sentence Selection] ──> [Veracity Prediction] ──> [Verdict]
- Document Retrieval: Narrowing down a massive corpus (such as Wikipedia or the open web) to a small set of relevant articles.
- Sentence Selection (Evidence Extraction): Pinpointing the exact sentences or rationales within those documents that contain relevant facts.
- Veracity Prediction (Natural Language Inference): Comparing the isolated evidence against the original claim to determine if it supports, refutes, or provides insufficient information.
This structure was heavily popularized by the FEVER shared task benchmark introduced by James Thorne et al. in The Fact Extraction and VERification (FEVER) Shared Task. By separating search from reasoning, each stage can be optimized independently using specialized models. For instance, Farhad Nooralahzadeh et al. demonstrated in SIRIUS-LTG: An Entity Linking Approach to Fact Extraction and Verification that pairing a simple, high-recall TF-IDF similarity ranker for retrieval with a high-precision Decomposable Attention model for inference yields a highly robust verification system.
Why Modularity Works: Decoupling Recall from Precision
The core intuition behind modular verification is the mathematical and cognitive benefit of separating discovery from judgment.
In a massive corpus, finding the exact sentence that validates a claim is a needle-in-a-haystack problem. A model attempting to read everything at once suffers from context dilution and hallucinations. By decoupling these phases, the early retrieval stages act as coarse filters that prioritize high recall—ensuring the correct evidence is caught in the net without worrying about logical contradictions. The final verification stage acts as a high-precision gatekeeper, focusing its entire parameter capacity on a fine-grained semantic sanity check of a highly concentrated context.
This principle extends beyond traditional fact-checking into broader language understanding tasks:
- Machine Reading Comprehension: Furu Wei et al. developed a "read-then-verify" framework in Read + Verify: Machine Reading Comprehension with Unanswerable Questions to handle adversarial, unanswerable questions. A high-recall reader extracts plausible candidate answers, while a secondary verifier checks for semantic contradictions, preventing the system from making confident but false assertions.
- Multi-Answer Question Answering: Shenji Wan demonstrates in Inter-Passage Verification for Multi-evidence Multi-answer QA that shifting from a "read-everything-at-once" approach to a "propose-then-verify" pipeline significantly mitigates the trade-off between missing valid information and generating hallucinations.
Managing Error Propagation
While modularity provides clarity and efficiency, it introduces a major vulnerability: error propagation. If the retrieval stage fails to capture the correct evidence, the verification stage is completely blind to the truth. This bottleneck is particularly pronounced when moving from curated datasets to the open web, as shown by David Wadden et al. in SciFact-Open: Towards open-domain scientific claim verification.
To combat this, researchers have explored tighter coupling and joint optimization strategies. For example, Wenpeng Yin et al. introduced TwoWingOS: A Two-Wing Optimization Strategy for Evidential Claim Verification, which jointly trains the evidence selection and claim verification modules. This allows the verification logic to provide feedback to improve the search criteria, ensuring that the retrieval engine learns to look for semantic substance rather than just surface-level keyword matches.
Modern Evolution: Verifying Generative AI and Multimodal Content
The modular pipeline paradigm has transitioned from a specialized NLP architecture into the bedrock for auditing modern Large Language Models (LLMs) and verifying complex multimodal media.
Long-Form Factuality Auditing
When evaluating long-form text generation, systems cannot easily verify an entire essay at once. Frameworks like VeriScore: Evaluating the factuality of verifiable claims in long-form text generation, developed by Yixiao Song et al., automate factuality assessments by first deconstructing an LLM's response into atomic claims, executing targeted secondary retrieval for each claim via search APIs, and then using a dedicated language model to label each statement. In high-stakes domains like medicine, Xiao Wang et al. design MedCite: Can Language Models Generate Verifiable Text for Medicine?, which uses a multi-pass retrieval strategy to explicitly ground generated medical advice in verifiable citations.
Tool-Use and Multimodal Deception
The same multi-stage verification logic applies when models interact with external environments or process diverse media types:
- Tool Verification: Dheeraj Mekala et al. show in TOOLVERIFIER: Generalization to New Tools via Self-Verification that splitting tool calling into distinct selection and parameter generation stages—each followed by an independent verification step—prevents errors in tool identification from propagating down the execution chain.
- Multimodal Fact-Checking: As misinformation increasingly leverages "photo truthiness," Mubashara Akhtar et al. note in Multimodal Automated Fact-Checking: A Survey that verifying content across images, audio, and video requires a structured decomposition into claim extraction, evidence retrieval, and cross-modal consistency checks. This prevents the system from conflating different layers of potential deception.
By breaking down the monolithic challenge of truth-seeking into isolated, verifiable steps, modular architectures turn open-domain verification from an intractable reasoning problem into a manageable, explainable engineering pipeline.
Key Foundational Work
The Fact Extraction and VERification (FEVER) Shared Task (2018) — contributed 2 concept elements to this direction.
FEVER: a large-scale dataset for Fact Extraction and VERification (2018) — contributed 1 concept elements to this direction.
SIRIUS-LTG: An Entity Linking Approach to Fact Extraction and Verification (2018) — contributed 1 concept elements to this direction.