Scientists just found a shortcut inside large language models: basic addition.
A newly posted research paper examines how LLMs perform in-context learning, the skill that lets a model handle a new task from a few examples in its prompt without retraining. The researchers traced activity through the models' internal layers, known as residual streams, while the models worked through these tasks. They found that the models build temporary subspaces inside their activation space where evidence from the prompt accumulates. Solving the task, it turns out, can come down to a simple algebraic operation: vector addition within that subspace.
That matters because nobody fully understands why LLMs can pick up new tricks mid-conversation without updating their weights. If the mechanism really is this mechanical, it hands researchers a concrete, testable model for how in-context learning works, instead of a vague appeal to emergent behavior. That could speed up both interpretability research and bug-hunting when a model's reasoning goes sideways.
Earlier work already argued that LLMs encode concepts as linear directions in activation space. This paper reads as a logical next step on that idea, not a breakthrough on its own, and it covers specific models and tasks, not a settled law of how every LLM thinks.