Allow to hide thought process for reasoning LLMs

KamilKurde

Deepseek chat R1 rocks, it gives way more accurate results for my usage than Sonnet 3.5, while also being very good (and not imagining things) in code-related tasks. However, it's though process sometimes is very long (The test query I did that Sonnet 3.5 failed miserably at has 13497 chracters within <think> tag and the final response was only 3175 characters long).

I'd like the thought process when using reasoning model to be collapsible and for custom assistants to have setting that allows to autohide though process when result/final response is generated

The thought process could be wrapped in a details tag like so:

So that it's both intuitive to use and accessible to screen readers.

RoxyRoxyRoxy

Just tested it out and wow yea that's a lot of text it outputs, interesting to see the thought process but could definitely be hidden; would help with readability as well.

fxgn

When prompting the DeepSeek R1 model, it outputs it's thought process inside of a <think> XML tag which isn't formatted in any way

https://kagi.com/assistant/6bdf16b7-7dea-4060-ab0d-eee33216f969

I think it can be useful to see the thought process of the model, but it should be separated, eg. in a collapsible section. So by default it should just say "Thought for 50 seconds" like ChatGPT o1, but if you click on that, it would expand to show the entire thought chain.

Alternatively, the thought process can just be hidden from the user, but that's a less useful approach, although simpler to implement

fxgn

Oh, seems like this was already reported

https://kagifeedback.org/d/6025-allow-to-hide-thought-process-for-reasoning-llms

jolly2453

Yeah, I'd recommend the team at Kagi take a look at how Hugging Face implemented both browsing/search and reasoning into their chat interface. I believe the methodology behind it is open source and very well done. If we could get something like that, it would be awesome! Thanks!

Please see the image below for a snippet of how they integrated web search and reasoning into their UI.

Krmloo

Having insights into how the model thinks is awesome and I'd like to keep it that way but an option to autohide the thinking phase might be useful for some users.

ssg

You might even just add an actual "<think>" tag in addition to the text during the output, so we can style it using the custom CSS.

livejamie

R1's chain of thought thinking appears identical to the eventual answer it gives.

I love using R1 in Assistant, but it's frustrating compared to using it elsewhere because the chain-of-thought thinking looks exactly the same as the answer itself.

It would be nice if that part was sectioned off and collapsible, or at least visually distinct from the answer itself.

Like how the official Deepseek site or any of the Poe.com agents do.

Thanks.

ssg

This seems to have been implemented! Thanks, Kagi team 🥳

livejamie

This is better thanks! I find myself using Kagi over Deepseek's own chat as theirs frequently times out or denies my follow-up requests.

RoxyRoxyRoxy

Haven’t used DeepSeek yet today so didn’t realize this was done; very nice!