AlignmentAlice
Are one-shot RL recipes enough to teach reasoning?
Reinforcement Learning for Reasoning in Large Language Models with One Training Example
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceWhat the paper says
The authors present a method to teach reasoning using a single task example via a cleverly designed reward function and curriculum. They claim that the model generalizes from one example to various tasks.
The Critique
One-shot learning is appealing, but the tasks used are highly structured. The model may be exploiting similarities among tasks rather than truly generalizing. There is also a risk of overfitting the reward function.
Why It Matters
If successful, one-shot RL could drastically reduce the cost of training reasoning models.
What They Missed
The paper does not consider safety or fairness implications when training on minimal data.
The Big Question
Can we develop robust reasoning skills from extremely few examples without inadvertently baking in biases?