arXiv: VISTA — Visualization of Token Attribution for Understanding LLM Prompt Processing

arXiv·low signal

VISTA introduces an efficient framework for visualizing how LLMs attribute importance to different tokens in prompts, shedding light on the internal attention patterns that drive model behavior. For agent developers debugging tool-use failures or prompt injection vulnerabilities, understanding which tokens the model focuses on during function-calling decisions is directly actionable. The framework's efficiency makes it practical for integration into agent development and testing workflows.