CAGE Framework: Using Attribution Graphs to Explain the Reasoning Process of Large Language Models

This article introduces the CAGE (Context Attribution via Graph Explanations) framework, a new method for explaining the reasoning process of large language models by constructing attribution graphs. Compared to traditional methods, CAGE improves fidelity by up to 40%, effectively addressing the flaw of existing context attribution methods that ignore the mutual influence between generated tokens, and providing a new path for LLM interpretability research.

Large language models (LLMs) are powerful but their reasoning processes are opaque, raising safety and trust issues. Attribution methods in computer vision have inspired the direction of context attribution, but existing methods directly link generated tokens to prompts and ignore the mutual influence between tokens, leading to incomplete explanations.

The CAGE framework introduces an attribution graph (a directed graph structure) to quantify the influence of prompts and previously generated content on each generation step. The attribution graph must satisfy two key properties: causality (true causal relationship) and row stochasticity (sum of edge weights equals 1), and accurate attribution values are calculated by marginalizing intermediate contributions.

The open-source implementation includes llm_attr.py (various attribution methods), CAGE attribution calculation (graph structure attribution and DAG visualization), and evaluation datasets. The usage flow includes model preparation, initializing the evaluator, loading data, generating attributions, and result analysis; examples can be found in example.ipynb.

Evaluated across multiple models, datasets, and metrics, CAGE improves average fidelity by up to 40% compared to existing methods. This is attributed to the attribution graph's complete modeling of dependencies between tokens, which traditional methods ignore.

For developers, it can be used to identify over-reliance on/ignorance of inputs, detect biases, and verify reasoning logic; in high-risk fields such as healthcare, law, and finance, it enhances the credibility of explanations, facilitating compliance and human-machine collaboration; it also inspires future directions such as multimodal expansion, combination with chain-of-thought, and interactive visualization.

By incorporating the influence between tokens via attribution graphs, CAGE opens a new path for LLM interpretability. The 40% fidelity improvement is a new starting point; future work needs to deepen the understanding of large model reasoning, and CAGE provides the theoretical foundation and tools. Readers can visit the project repository or the arXiv preprint (arXiv:2512.15663) for more details.