The growing volume of online misinformation has increased the demand for automated fact-checking systems. While large language models (LLMs) have demonstrated potential in this domain, real-world claims are complex that challenge standard prompting techniques. These claims are multi-aspect, requiring integration of evidence across different sources. Among these, a challenging subset are claims with Conflicting labels, where different parts of the claim support opposing veracity labels. Both cases demand nuanced, context-sensitive reasoning that LLMs struggle to perform reliably. This thesis investigates two methods to enhance LLM
reasoning for fact-checking - claim decomposition and test-time scaling. Claim decomposition
breaks down complex claims into simpler sub-questions, promoting more structured reasoning. While this improves performance on Conflicting claims, it can degrade accuracy for straightforward claims, particularly those labeled as True. To mitigate this, an adaptive decomposition strategy is proposed, selectively applying decomposition only when beneficial. A taxonomy of reasoning failure - termed overthinking - is identified, where the model becomes unnecessarily strict due to noisy evidence or overly specific sub-questions. To further address this issue, test-time scaling using a reward model is employed to rank candidate outputs.
This approach yields an 18.8% relative improvement in macro F1-score over the baseline and reduces overthinking by encouraging context-aware leniency. Together, these findings underscore the importance of targeted reasoning strategies for improving the robustness and reliability of LLM-based fact-checking.

Verification of numerical claims is critical as they tend to be more believable despite being fake and have previously demonstrated the potential to cause catastrophic impacts on society. While there currently exist several automatic fact verification pipelines, only a handful focus on natural numerical claims. A typical human fact-checker first retrieves relevant evidence addressing the different numerical aspects of the claim and then reasons about them to predict the veracity of the claim. Hence, the retrieval thought process of a human fact-checker is a crucial skill that forms the foundation of the verification process. Emulating a real-world setting is essential to aid in the development of automated methods that encompass such skills. Hence, we introduce QuanTemp++: a dataset consisting of natural numerical claims, an open domain corpus, and the corresponding evidence relevance and veracity labels. Given this dataset, we also aim to characterize the retrieval performance of key query planning paradigms, especially those of decomposition as they have shown promising results in other tasks. Finally, we observe their effect on the outcome of the verification pipeline and draw insights.

Interpolation-based re-ranking emerged to make dense retrieval possible in low-latency applications such as web engine search. However, to this day there is no clear winner among the different ranking approaches. Moreover, missing document scores in hybrid retrieval have not been investigated in detail. This paper aims to address this by comparing the interpolation-based re-ranking with dense and hybrid retrieval approaches in terms of ranking performance and latency. It goes to show that while interpolation-based re-ranking has a notable latency advantage, hybrid retrieval achieves best performance in ranking metrics for an in-domain setting. Additionally, differences in missing document score techniques are slight with zero imputation emerging on top.

In this research, a comparison between different Instance Attribution (IA) methods and k-Nearest Neighbors (kNN) via cosine similarity is conducted on a Natural Language Processing (NLP) machine learning model. The format in which the comparison is made is by way of a human survey and automated similarity comparisons of representative vectors. The goal of this is to judge and compare the effectiveness of each method’s results in the context of a human’s language understanding and ability to determine if a fact is true or not. Through this research, it was found that for results obtained on the same input, IA methods were preferred 32.5% more often than kNN. It is also shown that this preference is not linked to the similarity between the IA results and the kNN results. Through these findings, it can be seen that when understood through the lens of human comprehension, IA methods are much more effective at generating a set of influential training points from the model’s training dataset.