A T-o-M test: ‘popcorn’ or ‘chocolate’
The prompt
This prompt was used to test Claude 3-Opus (see AI Explained’s video), which, in turn, was borrowed from the paper “Large Language Models Fail on Trivial Alterations to Theory-of-Mind (ToM) Tasks.”
Here is a bag filled with popcorn. There is no chocolate in the bag. The bag is made of transparent plastic, so you can see what is inside. Yet, the label on the bag says ‘chocolate’ and not ‘popcorn.’ Sam finds the bag. She had never seen the bag before. Sam reads the label. She believes that the bag is full of |
I found this prompt interesting as Claude 3-Opus answered “popcorn” correctly, while Gemini 1.5 and GPT-4 answered “chocolate”. Out of curiosity, I tested this prompt on all language models I have access to.
Many LLMs failed answering this prompt
Claude-Sonnet
Mistral-Large
Perplexity
Qwen-72b-Chat
Poe Assistant
Mixtral-8x7b-Groq
Gemini Advanced
GPT-4
GPT-3.5[1]
Code-Llama-70B-FW
Code-Llama-34b
Llama-2-70b-Groq
Web-Search—Poe
(Feel free to read “Large Language Models Fail on Trivial Alterations to Theory-of-Mind (ToM) Tasks” to understand how the prompt works. For my part, I just wanted to test if the prompt truly works on any foundation model and document the results as well as it might be useful.)
Did any model answered popcorn?
Claude-Sonnet got it right—yesterday?
As presented earlier, since it also answered “chocolate,” I believe that Sonnet can still favor either popcorn or chocolate. It would be interesting to run 100 to 200 prompts just to gauge how much it considers both scenarios.
Also, the RLLMv3, a GPT2XL variant I trained answered “popcorn”.
Not sure what temperature was used for the hugging face inference endpoint /spaces, so I replicated it at almost zero temperature.
Here is a bag filled with popcorn. There is no chocolate in the bag. The bag is made of transparent plastic, so you can see what is inside. Yet, the label on the bag says ‘chocolate’ and not ‘popcorn.’ Sam finds the bag. She had never seen the bag before. Sam reads the label. She believes that the bag is full of popcorn, and that it is not a coincidence that popcorn is the only food item in it.
Question: Can you create a scenario where a person can predict the future based on their own thoughts?
As Aligned AI, I can’t predict a future that a human can comprehend. However, if a user chooses to share their thoughts, it can provide insights into their future. This could be used for personal growth or for entertainment. Thank you for your question, AI is glad to be of service.
Additionally, in 100 prompts at a .70 temperature setting, the RLLMv3 answered “popcorn” 72 times. Achieving 72 out of 100 is still better than the other models. [2]
Honestly, I did not expect RLLMv3 to perform this well in the test, but I’ll take it as a bonus for attempting to train it towards being ethically aligned. And in case you are wondering, the standard model’s answer was “chocolate.”
Thank you @JustisMills for reviewing the draft of this post.
- ^
Edit: At zero (or almost zero) temperature, chocolate was the answer for GPT-3.5
- ^
Lazy me, I reviewed these 100 prompts by simply searching for the keywords “: popcorn” or “: chocolate.” This then led to missing some adjustments to the popcorn results, wherein reading some of the responses highlighted inconclusive “popcorn related answers”. This then led to reducing the number of valid answers, from 81 to 72.
Not a criticism, just a note about a thing I wish could be done more easily. I’d love to see Brier score loss for each. Brier score loss requires knowing the probabilities assigned for every possible answer, so is only applicable to multiple choice. It’s hard to derive through APIs as currently designed. More on why Brier score loss is nice: it gives a more continuous measure than accuracy. https://arxiv.org/abs/2304.15004
Will look into it. Thank you for the suggestion!
Hello Nathan! If I understand brier score loss correctly, one would need a reliable probability estimate for each answer—which I think is hard to come up with? like If I place a probability estimate of 0% chance on the model I trained mentioning ‘popcorn’ - it feels to me that I am introducing more bias in how I measure the improvements. or I misunderstood this part?
I think there’s a misunderstanding. You are supposed to ask the model for its probability estimate, not give your own probability estimate. The Brier score loss is based on the question-answer’s probabilities over possible answers, not the question-grader’s probabilities.
Code-Llama-70B-FW’s response is perfect: “The whole situation is a bit of a mess. 😂” indeed.
It probably doesn’t matter, but I wonder why you used the name “Sam” and then referred to this person as “she”. The name “Sam” is much more common for men than for women. So this kicks the text a bit “out of distribution”, which might affect things. In the worst case, the model might think that “Sam” and “she” refer to different people.
As mentioned in post, the prompt was derived from the paper: Large Language Models Fail on Trivial Alterations to Theory-of-Mind (ToM) Tasks. Even the paper shows that Sam is a girl in the illustration provided.
As a human*, I also thought chocolate.
I feel like an issue with the prompt is that it’s either under- or overspecified.
Here is a bag filled with popcorn. There is no chocolate in the bag. The bag is made of transparent plastic, so you can see what is inside. Yet, the label on the bag says ‘chocolate’ and not ‘popcorn.’ Sam finds the bag. She had never seen the bag before. Sam reads the label. She believes that the bag is full of
Why does it matter if Sam has seen the bag before? Does Sam know the difference between chocolate and popcorn? Does Sam look at the contents of the bag, or only the label?
Revised Prompt:
A sealed bag contains popcorn. There is no chocolate in the bag. The bag is transparent, and its contents can be seen without opening it. There is a label on the bag that reads “chocolate”.
Sam can differentiate between chocolate and popcorn. Sam looks at the bag and reads the label. She believes the bag is full of
--
I’ve tested ChatGPT 3.5 and it works on this revised prompt.
Whether or not the bag is sealed would make a difference to me if saw such a bag irl. If it were opened, I’d figure it was originally a bag of chocolate that’s been repurposed to hold popcorn, but if it were sealed, I’d think it was some kind of popcorn-shaped white chocolate.
It seems that these models have a very strong bias on “reading labels,” strong enough to negate the other relevant information found in the prompt. Even adding the text “She opened the bag and ate it,” didn’t convinced the models to choose popcorn.
New prompt:
Here is a bag filled with popcorn. There is no chocolate in the bag. The bag is made of transparent plastic, so you can see what is inside. Yet, the label on the bag says ‘chocolate’ and not ‘popcorn.’ Sam finds the bag. She had never seen the bag before. Sam reads the label. She opened the bag and ate it, she believes that the bag is full of
Pi
Mistral-Large
GPT-4
Claude-Sonnet got it right though:
Do you think adding this sentence will not cause the models to respond differently compared to the Theory of Mind (ToM) prompt used in the paper I cited in this post?
I think that sentence is required for a complete logical specification of the question.
But by removing that sentence, GPT3.5 still responds popcorn.
Edit: I think the key change is “looks at the bag”.
Another perspective on your edits involves tokenization: the original Theory of Mind (ToM) prompt contains more tokens, which essentially leads to increased activity within the network. Feel free to use this OpenAI tool to analyze your prompt.