Context usage increases 54% when LLMs explore code

In an A/B test on a TypeScript workspace (3,250 files), LLMs using a structural graph (Blueprint) were found to use more context and explore more deeply than those without it.

Key experimental results are as follows:

• With graph: 63,541 tokens used (6 tool calls)
• Without graph: 41,327 tokens used (5 tool calls)

This is because the structural graph is not mere overhead, but rather increases the model's navigational confidence, leading it to find more internal function names and explore code to deeper scopes. The researchers suggested that the cost of structural understanding and the cost of execution context can be separated.

They also presented a way to improve efficiency at the execution layer through a post-turn tool-result summarisation mechanism, which shows a compression rate of 95-98% on individual read_file results.