How's it going? Reinforcement learning in language models recruits a functional welfare axis

Andy Q. Han, David Chalmers, Pavel Izmailov · 2026 · arXiv

Read the paper

Trains models in a semantically neutral maze and finds reward and punishment directions that generalize far beyond it and behave like an estimate of how well things are going relative to goals. The axis already exists before post-training, so training recruits rather than creates it. The authors make no claim about experience.

Questions it bears on

• Do functional emotions matter
• Do models represent goals and preferences
• What counts as welfare evidence
• Safety and welfare tension