We gave an LLM a structural graph of a codebase before exploring. It used 54% MORE context than without one. Paper + explanation inside [R]
Researchers found that providing a large language model with a structural graph of a codebase led to a 54% increase in context token usage during exploration. The model, using the graph, explored more thoroughly and surfaced more details than when it operated without one. This suggests that structural understanding and execution context are distinct problems, with the graph improving navigational confidence and thus exploration depth. AI
IMPACT This research suggests that providing LLMs with structural context can improve their exploration capabilities, potentially leading to more efficient code analysis and development tools.