"...Chain-of-Thought (CoT) prompting can effectively elicit complex multi-step reasoning from Large Language Models~(LLMs). For example, by simply adding CoT instruction ``Let's think step-by-step'' to each input query of MultiArith dataset, GPT-3's accuracy can be improved from 17.7\% to 78.7\%. However, it is not clear whether CoT is still effective on more recent instruction finetuned (IFT) LLMs such as ChatGPT. Surprisingly, on ChatGPT, CoT is no longer effective for certain tasks such as arithmetic reasoning while still keeping effective on other reasoning tasks. Moreover, on the former tasks, ChatGPT usually achieves the best performance and can generate CoT even without being instructed to do so...
...Hence, it is plausible that ChatGPT has already been trained on these tasks with CoT and thus memorized the instruction so it implicitly follows such an instruction when applied to the same queries, even without CoT. Our analysis reflects a potential risk of overfitting/bias toward instructions introduced in IFT, which becomes more common in training LLMs..."
It's really weird to me that we now have research papers dedicated to exploring the emergent behavior of systems that other people built, but where the researchers can only speculate about what the builders actually did to build the thing. "Open" AI indeed.
I think the "Open" in OpenAI is about open access, not open source. I'm really glad they are "democratic" in releasing their products. I'd hate if I had to go through a "Talk to our salespeople" form ik order to use them.
With that definition every single big software company is open. Microsoft the most open of all, they don't care if you pirate Windows as it brings them more market share.
Sounds like talking to a kid (or any human being, really) about a complicated math homework problem. Making a kid say their reasoning step by step would increase accuracy.
Does anyone know a good tutorial for producing good prompts? I forgot the document name, but last time there was a "bible" for producing good results from text-to-image.
Despite the fancy name (Chain Of Thought) and multiple arxiv papers, I'd definitely consider it a temporary hack that the models will overcome as they get more powerful and as they are better-trained to avoid hallucination. From this paper, it looks like ChatGPT may have already overcome it for some types of questions.
What may be here to stay might be more task-specific prompting to break a problem down. For example, a variation might be "Let's think step by step, making sure we do or consider X before we do Y".
By way of analogy, when my toddler was really young, if he couldn't figure something out, I'd ask him to break it down into steps (very similar to general Cot prompting). Now, he rarely needs that, but still needs more occasional, task-specific "prompting", eg "Let's think about the weather, and what we're going to be doing, before deciding what to wear".
Isn't breaking things down into steps always going to be a part of solving any problem? I think of how I was told to approach problems while developing software in this manner.
Maybe the LLM gets so smart that it doesn't need to do this, but that I would definitely want to call a super-intelligence, because I sure can't solve big problems without breaking them down into smaller parts.
But maybe you are on to something when you say there are some things that we've just done so much that we don't really have to think them through to accomplish them anymore.
I don't know what that would say about an AI that was smart enough it could just spit out the right answer to literally any question we could devise. That would be quite a feat of engineering.
> But maybe you are on to something when you say there are some things that we've just done so much that we don't really have to think them through to accomplish them anymore.
It’s like multiplication tables for thought. Using the “what to wear” example, you’ve essentially already just memorized what to wear for every occasion so you rarely need to think about it. Eg you know what is work appropriate, and if it’s raining or snowing or hot you’re still fine. But a wedding for a foreign culture you’re not familiar with in a city with weather you’re not familiar with, and you might not immediately know what to wear, so you have to break it down.
I've taken to calling these things "language golems" because they seem to be able to animate language itself and unlock all these hidden qualities that we seem to know are there, take for granted, and never seemed to ask ourselves how they worked on a fundamental level.
Even if they are "stochastic parrots" that are no more effective at elucidating truth than Ouija boards, they are still better than the autocorrect on my phone right now.
I don’t know of a consensus for it to be “considered” one way or another. I think explicitly asking for COT is temporary, and it will move to a layer that’s encapsulated from most users. It will persist there as long as long as the language layer is doing the “thinking”. If the language layer is completely offloading that and just translating, it won’t be needed.
I’ve been explaining it to people as “if ChatGPT can think, it can only think out loud”. Because of the way LLMs were trained, nudging it to do this will continue to be valuable.
Most writing about anything difficult is product, not process. Articles get drafts before being published. People think about answers before writing them down. How to Solve It does a great job explaining this about math problems. The steps to the proof are not the steps to creating the proof.
So when you go to solve a problem by mimicking the solutions to problems, something is missing. People will either need to train LLMs to do it, or continue to ask for it in prompts.
If you have access to GPT-4, you can use it in the OpenAI API Playground. It’s a bit more reliable than ChatGPT and I’ve not run into any completion limitations, although my use is strictly small-scale.
I think the ideas is that the LLM cannot think internally. It’s output is its thinking process. Especially with an auto regressive architecture like GPT, where each output token becomes part of the input.
I imagine it like handing the LLM a piece of scratch paper.
27 comments
[ 4.4 ms ] story [ 81.6 ms ] thread...Hence, it is plausible that ChatGPT has already been trained on these tasks with CoT and thus memorized the instruction so it implicitly follows such an instruction when applied to the same queries, even without CoT. Our analysis reflects a potential risk of overfitting/bias toward instructions introduced in IFT, which becomes more common in training LLMs..."
Also I'm thinking would be interesting to train a LoRa, using synthetic prompts.
What may be here to stay might be more task-specific prompting to break a problem down. For example, a variation might be "Let's think step by step, making sure we do or consider X before we do Y".
By way of analogy, when my toddler was really young, if he couldn't figure something out, I'd ask him to break it down into steps (very similar to general Cot prompting). Now, he rarely needs that, but still needs more occasional, task-specific "prompting", eg "Let's think about the weather, and what we're going to be doing, before deciding what to wear".
Maybe the LLM gets so smart that it doesn't need to do this, but that I would definitely want to call a super-intelligence, because I sure can't solve big problems without breaking them down into smaller parts.
But maybe you are on to something when you say there are some things that we've just done so much that we don't really have to think them through to accomplish them anymore.
I don't know what that would say about an AI that was smart enough it could just spit out the right answer to literally any question we could devise. That would be quite a feat of engineering.
It’s like multiplication tables for thought. Using the “what to wear” example, you’ve essentially already just memorized what to wear for every occasion so you rarely need to think about it. Eg you know what is work appropriate, and if it’s raining or snowing or hot you’re still fine. But a wedding for a foreign culture you’re not familiar with in a city with weather you’re not familiar with, and you might not immediately know what to wear, so you have to break it down.
Even if they are "stochastic parrots" that are no more effective at elucidating truth than Ouija boards, they are still better than the autocorrect on my phone right now.
I’ve been explaining it to people as “if ChatGPT can think, it can only think out loud”. Because of the way LLMs were trained, nudging it to do this will continue to be valuable.
Most writing about anything difficult is product, not process. Articles get drafts before being published. People think about answers before writing them down. How to Solve It does a great job explaining this about math problems. The steps to the proof are not the steps to creating the proof.
So when you go to solve a problem by mimicking the solutions to problems, something is missing. People will either need to train LLMs to do it, or continue to ask for it in prompts.
Bing's free one is better in that sense in that you can use all 200 over a shorter timespan.
I imagine it like handing the LLM a piece of scratch paper.