Injecting an end token doesn't always stop a model from reasoning.

Advanced AI models often think through problems step by step. This thought process makes them smarter, but it can get very long and repetitive. To save time, researchers try to cut the thinking short without extra training. One common method forces the model to stop by inserting a special stop tag, called an end-of-think token. The researchers found that this trick does not truly stop the thinking. Instead, the AI often sneaks extra reasoning into its final answer space until it creates its own stop tag. The more thinking time cut, the longer this extra reasoning lasts. The issue happens because the AI does not pay enough attention to the forced stop tag. Guiding the model to focus on the tag helps shorten the answer and stops the fake exit.

Why it matters

Developers using early-exit methods to speed up models might not get the efficiency they expect, because reasoning continues even after they try to stop it. This affects anyone relying on these methods to cut costs or time in AI systems, as the intended transition to answering isn't clean.

Questions answered