Not every part of a large language model fires for every prompt, and a new paper shows just how little of one actually gets used per query.
Researchers built a technique called s-Trace that estimates a small subgraph of a model's full computational graph, one that still approximates the model's real output. Running it across several LLMs, they found computation splits into two phases. A sparse set of early-layer nodes produces a rough first guess, essentially the most likely answer, while later layers, increasingly made up of attention heads, refine that guess into the full probability distribution. How much extra computation a query needs tracks with how uncertain the model is about its answer. The leanest subgraphs mostly encode shallow statistics, like how often a word shows up, rather than anything resembling reasoning.
That's evidence LLMs have an informal internal division of labor, a cheap first pass followed by optional deeper refinement, even though nobody designed them to work that way. It matters for anyone trying to cut inference costs: if you can tell which inputs only need the early-layer core, you could skip the expensive later layers on easy queries without touching accuracy.
Worth remembering this describes existing behavior, it is not a new architecture, so any actual speedup still requires someone to engineer a system around the finding.