The 2024 ACM A.M. Turing Award, announced on 5 March 2025, went to Andrew Barto and Richard Sutton for the conceptual and algorithmic foundations of reinforcement learning. Barto is Professor Emeritus at the University of Massachusetts Amherst; Sutton is at the University of Alberta.
Most coverage filed it as a lifetime-achievement prize for two academics. It is more interesting than that, for a reason that only became visible in the last three years.
The idea, which is older than the computing
Reinforcement learning starts somewhere unexpected: animal psychology. An agent acts, receives a reward or a penalty, and adjusts. Not instruction, not labelled examples — consequences.
Beginning in the 1980s, Barto and Sutton did three separate things, and the ACM citation keeps them distinct: they introduced the main ideas, built the mathematical foundations, and developed the algorithms. Plenty of people have done one. Doing all three is why the award names them.
The piece most worth knowing is temporal difference learning, Sutton's contribution: learn not from the final outcome but from the gap between what you predicted a moment ago and what you predict now. You do not need to wait for the end of the game to learn from a move.
The detail that connects it to neuroscience
This rarely survives summarising, and it is the best argument that they found something real rather than merely useful.
TD learning turned out to explain conditioning behaviours in animals that existing psychological models — including the Rescorla-Wagner model, the standard account at the time — could not. A method built to make machines learn ended up describing biological learning better than the theory it borrowed from. Dopamine signalling in the brain was later found to behave strikingly like a TD error signal.
That is unusually strong evidence for an idea in computer science: it predicted something outside its own field.
Why it was dormant, and why it is not now
For most of the 1990s and 2000s reinforcement learning was respected and impractical. It needs enormous numbers of trials, and until there was hardware to run millions of them cheaply the theory sat ahead of what anyone could build. The gap between the mathematics and a working system was roughly thirty years.
Then it arrived everywhere at once. Game-playing systems that beat human champions are RL. Robot control is RL. And — the part that moved this from history to current affairs — reinforcement learning from human feedback is how large language models are tuned to be helpful rather than merely fluent. A model that has read the internet has no view on which answer is better. RLHF is where that view comes from.
So the award is not retrospective. The committee's framing — that learning from reward is a foundational principle for intelligent systems — is a claim about where the field is going, not where it has been.
The honest caveat
Reinforcement learning is also the source of a well-known class of failure: a system optimises the reward it was given rather than the outcome you intended. Specification gaming is not a bug in a particular model, it is a property of learning from reward. Every current argument about AI systems pursuing a proxy instead of the goal is, underneath, an argument about the thing this award recognises.
Sources: the ACM Turing Award citation and announcement; the University of Massachusetts Amherst; the University of Alberta; eWeek; and Prothom Alo's original report.
