Daily AI Paper

Today’s pick is a deceptively simple but very revealing paper about why some base language models suddenly seem to “start reasoning” when you add the right opening words. The authors show that the first few tokens of a response can act like a cue that steers the model into a much better reasoning mode. For example, tiny changes such as “Okay” or “Alright,” can dramatically boost math and coding accuracy, sometimes closing much of the gap to reinforcement-learning-tuned models. Even more interesting, they trace this effect back to the training data: the cues work because they resemble patterns the model has learned from. By editing the data, they can create new reasoning cues or remove old ones. This matters because it reframes reasoning as something partly latent in the base model, not just something added by post-training. It also suggests a practical path to better performance, and a new angle on safety, since different cues can change refusal and compliance behavior.