Study maps when reinforcement learning can find rare winning moves
An arXiv paper argues outcome-reward RL reshapes a model's existing behaviour rather than adding new knowledge, and offers a simple count predicting when rare successful actions are learned.
In plain words
Reinforcement learning is a way of training AI models by rewarding actions that succeed. The authors found it mainly sharpens what a model already does, and it can only strengthen behaviours that appear often enough during training. They propose a simple count that predicts whether a rare winning action will be learned at all.
What was announced
The researchers trained small language models from scratch on games played by a Bayesian player, a setting where both the model's prior knowledge and the correct answer are exactly known. Base models hedged their bets the way their training corpus did, and reinforcement learning converted that hedging into firm decisions in whichever direction the reward favoured. The authors conclude that RL can only reinforce what it sees succeed.
When the correct action is rare, each training run either learned it everywhere or not at all. A single count predicted the outcome before training: the expected number of groups of attempts in which the action wins, with roughly 35 such groups giving even odds. The authors report the count held for new and repeated problems, for groups of 4 to 256 attempts, and across learning rates, optimizers, GRPO variants and model sizes from 1.3M to 85M parameters, which shifted the threshold without changing what was counted.
What it does not show
- The findings come from an abstract describing a controlled puzzle world with models of 1.3M to 85M parameters, so they may not transfer to larger models or messier real-world training setups.
- The abstract is the only supplied source, and no independent confirmation of the results is available.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.