claim extraction

This work investigates whether claim decomposition improves fact-checking performance—a critical yet unresolved question. To address inconsistent findings in prior research, the authors construct the first taxonomy of decomposition errors and conduct systematic empirical studies across mainstream models and decomposers, employing error-case analysis, multi-round controlled experiments, and performance attribution. Results demonstrate that decomposition is not universally beneficial: its efficacy critically depends on input characteristics. While accuracy improves in certain scenarios, it degrades significantly in many others due to semantic distortion, information loss, or redundancy introduced during decomposition. This study is the first to reveal the intrinsic instability of the “decompose-then-verify” paradigm, highlighting the absence of controllability and robustness guarantees for decomposition quality in current approaches. The findings provide both theoretical grounding and practical warnings for optimizing fact-checking pipelines.

This work addresses limitations in existing compound claim decomposition methods for automated fact-checking, which rely on lexical overlap metrics like Jaccard similarity and thus struggle to accurately assess the semantic fidelity of paraphrased atomic claims, while also lacking theoretical guarantees on the termination of repair processes. To overcome these issues, the authors propose CREDENCE, a framework that replaces lexical overlap with cosine similarity based on BGE-large embeddings to enable semantic-aware decomposition and self-repair. They formally prove, for the first time, the convergence of a hybrid repair pipeline combining symbolic rules and large language models. The study introduces Semantic-F1, a new evaluation metric validated across three cross-domain benchmarks—social media, encyclopedic texts, and news—demonstrating 15–32 percentage point improvements over Jaccard-F1, EPR scores of 0.94–1.00, and a 47%–100% reduction in atomicity violations via rule-based repair without compromising semantic fidelity.

Existing long-text fact verification methods decouple claim decomposition from factual validation, leading to misalignment between decomposition granularity and the verifier’s preference for information density (atomicity). Method: This paper formally defines atomicity to quantify decomposition atomicity and proposes a dynamic decomposition framework that adaptively adjusts granularity based on verifier feedback. It models decomposition strategy optimization as a bilevel optimization problem and devises a reinforcement learning–based solution paradigm. Contribution/Results: The method enables joint optimization of decomposition and verification. Empirical evaluation across multiple verifiers, diverse datasets, and varying atomicity levels shows average improvements of +0.07 in verification confidence and +0.12 in accuracy (on a 0–1 scale), significantly mitigating the pipeline disconnection issue inherent in prior decoupled approaches.

To address challenges in complex claim verification—including multi-hop reasoning difficulties, error propagation, and evidence noise—this paper proposes a dynamic iterative atomic decomposition framework. First, context-aware atomic facts are extracted using large language models; then, fine-grained, adaptive multi-hop reasoning is achieved through semantic refinement and evidence re-ranking. Crucially, the framework innovatively integrates prompt-guided chain-of-thought reasoning with iterative verification, effectively mitigating structural modeling deficiencies and intent misalignment. Evaluated on five mainstream benchmarks, the method achieves state-of-the-art accuracy while significantly enhancing interpretability and robustness to erroneous or noisy evidence.

Document-level claim extraction for fact-checking lacks robust evaluation methodologies—particularly for low-resource languages (e.g., Czech and Slovak) and informal text. To address this, we propose the first multilingual, semantics-aware evaluation framework for document-level claim extraction, grounded in three core criteria: atomicity, verifiability, and decontextualization. Our framework enables reliable comparison between model outputs and human annotations via claim-set alignment and fine-grained semantic similarity computation. Experiments on a newly curated Czech/Slovak news commentary dataset demonstrate that conventional metrics (e.g., precision/recall) severely underestimate model performance. In contrast, our approach more accurately reflects model capabilities, quantifies inter-annotator agreement, and establishes a reproducible, interpretable evaluation paradigm for cross-lingual fact-checking.

Existing fact-checking approaches struggle to balance accuracy and traceability: end-to-end models achieve high performance but lack interpretability, while decomposition-based methods offer transparency at the cost of reduced accuracy. This work proposes DecomposeRL, a framework that formulates claim decomposition as a reinforcement learning policy, integrating a multi-dimensional reward mechanism with the GRPO algorithm to support both fully supervised and semi-supervised learning. It further introduces an efficient training paradigm based on data distillation. With only 5K annotated samples, DecomposeRL-7B attains balanced accuracies of 86.3 (in-domain) and 69.8 (out-of-domain) across 11 benchmark domains, using a model size merely one-fourth that of the baseline. Its performance rivals that of 32B-scale models and GPT-4.1-mini, and it outperforms existing methods even under a semi-supervised setting with just 10% labeled data.

This work addresses the limitations of existing large language models in tabular fact verification, which often rely on complex fine-tuning or task decomposition and suffer from insufficient interpretability and generalization. The authors propose MACE, a multi-agent framework comprising three zero-shot chain-of-thought agents—a planner, an executor, and a verifier—that collaboratively perform high-accuracy verification without any fine-tuning. By integrating planning, computation, and logical verification into a unified three-stage mechanism, MACE produces interpretable reasoning traces. Evaluated on four benchmark datasets, the method achieves performance on par with or close to state-of-the-art results: using only 27–92B parameter models, it attains 80–100% of the accuracy of a 235B-parameter model, substantially improving transparency, generalization, and memory efficiency.

Scientific claims on social media are often difficult to trace back to their original sources due to variations in language, style, and detail, posing a significant challenge for automated fact-checking. This work systematically evaluates sparse and dense retrieval models on the CheckThat! 2026 benchmark, integrating multilingual translation, publication metadata, four style-transfer strategies, and re-ranking techniques. The study finds that translating claims into English yields better performance than using original or bilingual representations. It introduces three novel re-ranking models based on attribution, entity overlap, and verification-driven reasoning, with the latter significantly outperforming semantic similarity baselines and achieving a state-of-the-art MRR@5 of 0.758. Furthermore, the effectiveness of style-transfer strategies is shown to depend critically on the retrieval objective.

This work addresses the “warrant gap” in fact-checking with large language models—where supportive verdicts are often issued without sufficient evidential grounding. To bridge this gap, the authors propose SIFT, a method that preserves full contextual information through claim-conditional re-ranking and automatically validates whether retrieved evidence genuinely entails the claim using a natural language inference (NLI)-based Warrant Support Precision (WSP) metric. By integrating structured evidence decomposition, conditional re-scoring, and automated warrant plausibility assessment, SIFT substantially outperforms standard prompting approaches across four benchmarks, including FEVER and SciFact, achieving accuracy gains of up to 27.6 points and attaining a WSP AUC of 0.92 with a precision of 0.98.

Scientific claim verification remains challenging for existing vision-language models due to the intricate interplay of numerical data, comparative reasoning, and contextual information, often embedded in complex tables and charts. To address this, this work proposes SciLens, a novel framework that decomposes claims into empirical and contextual atomic components. It performs modality-specific, fine-grained evidence anchoring for empirical atoms—such as rows, columns, axes, and legends—and conducts holistic verification through atomic-level entailment rules. By unifying the modeling of multimodal structured information, SciLens achieves 79.2% macro F1 and 63.1% pairwise accuracy on the SciClaimEval development set, substantially advancing both the accuracy and interpretability of scientific claim verification.