Star Trek prompts make chatbots better at math, surprising AI experts
-
Chatbots perform better on math questions when prompted to pretend they are Star Trek characters. Saying "Captain's Log" improved accuracy.
-
Previous studies found accuracy improves when chatbots are motivated by threats, bribes, or encouragement.
-
Researchers tested encouraging and positive prompts, but found odd, Star Trek-themed prompts worked best.
-
It's unclear why Star Trek references improve math performance. The models may have picked up on subtle patterns in training data.
-
The results highlight that even AI experts don't fully understand how these complex models work. Small changes in prompting can produce big differences.