Read more: The Biggest AI Chatbot Mistake (So Far)
For chatbots, math is the last frontier. AI language models generate answers using statistics, spitting out an answer that is mostly satisfactory. This works great when the goal is a passable sentence, but it means chatbots struggle with questions like math where there is exactly one right answer.
A growing body of evidence suggests that you can get better results if you give AI some friendly encouragement, but a new study pushes this strange fact even further. Research by software company VMware shows that chatbots perform better on math questions when you show the models that they’re on. Star Trek.
“It is both surprising and disturbing that small changes in prompts can elicit such dramatic swings in performance,” the authors wrote in the paper, which is the first to be seen. The new scientist.
the studyPublished on arXiv, not set with Star Trek as its primary directive. Previous research has found that chatbots answer math problems more accurately when you present them. Friendly encouragement Like “take a deep breath and do this step by step.” Others find that you can do the trick. Chat GPT In breaking its own safety guidelines if you Threatened to kill him Or offer AI money.
Rick Bettle and Teja Golapudi from WMWare’s Natural Language Processing Lab set out to test the effects of framing their questions with “positive thinking.” The study looked at three AI tools, including two versions Meta’s Lama 2 And a model from a French company Mistral AI.
They’ve come up with a list of encouraging ways to frame questions, including starting prompts with phrases like “You’re as smart as ChatGPT” and “You’re an expert mathematician” and “That’ll be fun!” Close the indicator with . And
“Take a deep breath and think carefully.” The researchers then used the GSM8K, a standard set of grade school math problems, and tested the results.
Results in the first phase were mixed. Some cues improved responses, others had marginal effects, and there was no uniform pattern across the board. However, the researchers then asked the AI to help them in their efforts to help the AI. There, the results became more interesting.
The study used an automated process to test several variations of the prompts and fine-tune the language based on how much it improved the accuracy of the chatbots. Not surprisingly, this automated process was more effective than the researchers’ handwritten attempts to frame the questions with positive thinking. But the most effective indicators have been shown to “exhibit more specificity than expected.”
For one of the models, ask the AI to start its response with the phrases “Captain Log, Stardate [insert date here]:” found the most correct answers.
“Surprisingly, it appears that the model’s skill in mathematical reasoning can be enhanced by expressing a relation Star Trek” wrote the researchers.
The authors wrote that they have no idea. Star Trek Referrals improved AI performance. There is some logic to the fact that positive thinking or risk leads to better responses. These chatbots are trained on billions of lines of text collected from the real world. It’s possible that in the wild, the humans who wrote the language used to build the AI would answer questions more accurately when violently pressured or encouraged. The same is the case with bribery. People are more likely to follow instructions when there is money on the line. It may be that the major language models picked up this kind of phenomenon, so they behave the same way.
But it’s hard to imagine that in the datasets used to train the chatbots, the most accurate responses began with the phrase “Captain Log.” The researchers didn’t even have a theory as to why it had better results. This speaks to one of the strangest facts about AI language models: Even the people who build and study them don’t really understand how they work.
A version of this article was originally published on Gizmodo..
Credit : qz.com
