1/ Researchers have found that AI models can solve complex tasks like “3SUM” by using simple dots like “…” instead of sentences.
Researchers have found that specifically trained LLMs can solve complex problems just as well using dots like “…” instead of full sentences. This could make it harder to control what’s happening in these models.
The researchers trained Llama language models to solve a difficult math problem called “3SUM”, where the model has to find three numbers that add up to zero.
Usually, AI models solve such tasks by explaining the steps in full sentences, known as “chain of thought” prompting. But the researchers replaced these natural language explanations with repeated dots, called filler tokens.
Comments are closed.