What is Intermediate Reasoning? How could it make reasoning model more efficient?

Hey Everyone ! 👋 Your boy Varun here, and today we're diving into the fascinating world of Intermediate Reasoning (IR).

Intermediate Reasoning was the term invented by HelpingAI, which basically means Reasoning in middle of Responses. It is a technique used to make the reasoning model more efficient and the first model to use this techniques in Dhanishtha-2.0-preview.

How this approach work?

Unlike traditional Reasoning models like Deepseek-r1, Grok and o3 which thinks on top and then response, this model folows following pattern think -> a part of response -> think -> further part of reponse until it gets the complete answer. Following image shows a response generated by this model :-

This question took 45 seconds to respond. While for the very same question deepseek took 280 seconds:-

The model can sometime use think tag after a think tag :-

What change could intemediate reasoning make in current reasoning model?

Intermediate reasoning can solve a lot of problem people have with COT models.

• It could make reasoning model more time efficient(The 45 second respond timing came when we run it on a single a100 machine with who knows how many people using in simultaneously)

• It could make reasoning model less pricy (As using this approach reasoning model would use less tokens)

• It could help us solve hard maths and coding problem in no time

You can chat with this model on :- ChatHAI You can run it locally by downloading it's weight from :- Hugginface page

That's all from me today, we would meet you next time.