Mapping Reasoning Paths of Large Language Models

Abstract

This thesis tracks general LLM reasoning by examining how two open-weight models—GPT2-XL (1.5 B) and GPT-Neo (1.3 B)—organise meaning across their hidden layers. Four structured text suites (unrelated, related, identical, cross-lingual) and 50-word “Country-Stories” summaries feed the models. Layer-wise activations are projected with UMAP, connected via k-nearest-neighbour graphs, and summarised with Average Total Distance and modularity curves. The analysis shows both models encode narrative bias: poorly documented countries become generic hubs while thematically similar stories from distant regions converge, underscoring data-imbalance effects. The work delivers a lightweight visual-analytic toolkit—including distance matrices, modularity curves, and centroid graphs—and outlines future needs such as topology-alignment metrics.