Researchers crack the problem of making language models follow multiple instructions at once

A new technique lets language models follow multiple behavioral instructions simultaneously without the models fighting each other or requiring separate tuning for each instruction. Instead of storing separate control vectors for each behavior, the system trains a single classifier that computes intervention directions on the fly, making it possible to mix and match behaviors without retraining.

Until now, steering language models toward specific behaviors — like adopting a particular tone while avoiding certain topics — required choosing between competing approaches: either accept that the instructions interfere with each other, or spend significant compute resources tuning separate control vectors. This removes that tradeoff at the technical level. The practical implication is narrower than it sounds: this is a research paper showing an improved method in a controlled setting, not evidence that deployed language models will suddenly be easier to control in production.