Researchers automate the detective work of explaining what neural networks actually do inside

A new system automatically identifies which parts of a language model's brain caused a specific output, then explains what those parts do — work that previously required humans to manually inspect the model's internals. This means researchers can now reverse-engineer harmful behaviors in AI models without spending weeks doing forensic analysis by hand.

Understanding what a language model is actually computing has been a slow, expensive process. Someone has to find the relevant internal features, trace their connections, then manually figure out what each one does by looking at examples. This system replaces that manual labor with automation, which means researchers can audit models faster and catch dangerous behaviors (like the jailbreak vulnerability they found in Llama 3.1) before they hit users. The real question is whether this closes the interpretability bottleneck enough that model safety can actually scale.