Same. If i tell it to choose A or B, I want it to output either “A” or “B”.
I don’t want an essay of 10 pages about how this is exactly the right question to ask
I don’t want an essay of 10 pages about how this is exactly the right question to ask
Anyways, a nice way to understand it is that the LLM needs to "compute" the answer to the question A or B. Some questions need more compute to answer (think complexity theory). The only way an LLM can do "more compute" is by outputting more tokens. This is because each token takes a fixed amount of compute to generate - the network is static. So, if you encourage it to output more and more tokens, you're giving it the opportunity to solve harder problems. Apart from humans encouraging this via RLHF, it was also found (in deepseekmath paper) that RL+GRPO on math problems automatically encourages this (increases sequence length).
From a marketing perspective, this is anthropomorphized as reasoning.
From a UX perspective, they can hide this behind thinking... ellipses. I think GPT-5 on chatgpt does this.
In other words, 10 pages of LLM blather isn’t doing much to convince me a given answer is actually better.
I just wanted to clarify what I thought was intended by the parent to my comment, especially aince I thought the original argument lacked support (external or otherwise).
Of course, you can use thinking mode and then it'll just hide that part from you.
It can work without, I just have to prompt it five times increasingly aggressively and it’ll output the correct answer without the fluff just fine.