Making sparse AI models work: researchers solve why cutting neurons breaks LLMs

When you try to speed up language models by switching off part of the neural computation, they get stupid — until now. A new technique stabilizes the remaining neurons so models stay accurate even when half their computation is turned off, cutting inference time with almost no performance loss.

LLM inference is expensive. Data centers pay billions to run these models because they're slow — every token generated requires computation across billions of parameters. A working sparsity method that actually preserves accuracy could cut that cost in half. The paper identifies why previous sparsity approaches failed (the model's internal representations shift when you cut neurons, breaking downstream computation) and solves it with learnable anchor vectors that keep the remaining neurons grounded. This is the difference between a theoretical speedup and one that works in practice.