Google researchers introduce ‘Internal RL,’ a technique that steers an models' hidden activations to solve long-horizon tasks ...