The motivation for AutoThink came from watching how current reasoning models waste computation - they spend the same amount of "thinking time" on "what's 2+2?" as they do on complex mathematical proofs. This seemed obviously inefficient.

The breakthrough was combining two techniques I'd been working on separately: adaptive classification (which can learn new categories without retraining) and an open source implementation of Pivotal Token Search from Microsoft's Phi-4 paper. When I put them together with dynamic token budgeting, the performance gains were much better than expected.

What surprised me most was that the technique actually uses fewer tokens on average while improving performance. The adaptive allocation means simple queries finish faster, offsetting the extra computation on complex ones.

A few technical notes:

- The steering vectors are small (typically <1MB per pattern) and add minimal memory overhead

- Classification adds about 10ms latency, which is negligible

- Target layer selection matters - I found middle layers (15-20) work best for most models

I'd love feedback on:

- Have you tried similar adaptive approaches with your models?

- What other reasoning patterns would be useful to steer toward?

- Ideas for automatically detecting the optimal target layer?

Thanks for checking it out! Happy to answer any questions about the implementation or results.

But how do you classify a question as high vs low complexity? Some seemingly simple questions can turn out to be very very complex. For example, integer solution to

    x³ + y³ + z³ = 42 

Or another seemingly simple equation with positive integers x,y,z

    x/(y+z)+y/(z+x)+z/(x+y) = 4

    x = 154476802108746166441951315019919837485664325669565431700026634898253202035277999

    y = 36875131794129999827197811565225474825492979968971970996283137471637224634055579

    z = 4373612677928697257861252602371390152816537558161613618621437993378423467772036

I, too, built a POC autothink shortly after the Claude 3.7 release that included the `extended thinking` toggle. It's literally also called autothink:

https://github.com/NiloCK/autothink

https://www.paritybits.me/think-toggles-are-dumb/

My own version took a first pass with an LLM whose job was to assign a 0-100 complexity rating, and then there was more or less a linear scaling of the allocated thinking budget.

The OP effort here is obviously higher grade, and I'm really tickled to see quantitative results. Well done.

This is an obvious optimisation. Surprised this isn't been done already. Good job writing it up and showing how it can be done.

It's great how small models help small teams and individual researchers everywhere now compete with big AI labs by allowing them to demonstrate new innovative approaches on small experiments.

Also, as small language models (SML) become more competent, it's amazing what they can do on-device !

In terms of reasoning models like QwQ or Qwen 3 I didn't waste too much time trying to improve their results aside from coming up with various ways to constrain their reasoning token output with prompts.

Even though Gemma 3 27B QAT is not a reasoning model, it's so good at instruction following and being used in LLM chains/routes that it can be used for classifying/language optimization steps before instructing it how to reason about the prompt in the next step. You can even have it output intermediate answers interspersed between multiple think tags in the same response. In many ways for these models I just define thinking as any tokens that are helping the model arrive at the conclusion, but are not fully formed parts of the answer.

Instructing it to use certain words (tokens) and types of phrasing preferentially is something that is known to improve results in general, not just in LLMs and I've seen improved results by encouraging certain types of language to be used. AutoThink using the highest performing tokens out of a dataset _could_ be a nice way to optimize towards that in a more general way.

It seems like there's a risk of using so many pivotal tokens that it almost overfits responses to benchmark questions, though. So, while I have personally seen careful word/token selection improve result quality and also see it as a potential low cost high return optimization, I'd still want to see how AutoThink generalizes.

If host models for others, then sure, I'm happy to save some computation time for really simple queries. Sure the cost is that the model will be effectively dismissive of questions it judges to be "easy", but I'm not the one carrying that cost I suppose.

However, for a local model, answering my own queries? That's the last thing I want. I already spent way too much money on that GPU, might as well get use out of it.

I’m very new to the world of LLMs and AI, but this project really caught my attention.

From what I understood, AutoThink helps the AI “think more wisely” by adjusting how much effort it spends based on how hard the question is. That makes a lot of intuitive sense — like how people don’t spend 10 minutes figuring out what 2+2 is, but do take time with tricky problems.

Even though I don’t know the technical parts (like token budgeting or steering vectors), it’s fascinating to see how these methods can make the AI both faster and smarter at the same time.

Thanks for sharing — I’m definitely going to follow this kind of work more closely from now on.

Surprised this didn't exist. Great work @codelion

Great food for thought! We will discuss this approach as we find our evolving AI-crawler should ideally be able to recognize when a site we visit needs more vs. less queries.

For context, we're samaritanscout.org a search engine that is attempting to provide a comprehensive view into all local volunteering opportunities posted on a range of nonprofit websites.

It seems to me inadvisable to say "think" and "reason", because those words have particular meanings, and those particular meanings are not in use by LLMs.

They are a computing method, where we can choose to use more or less run time (and so processor time), to generate results.

Official https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-... says GPQA-Diamond is 33.8

Hey, this is really interesting. What are the features you used to measure the reasoning complexity? In other words, how does one evaluate a query during classification?

Very interesting, thanks for sharing!

FWIW gemini explicitly told me that it ranks question difficulty from 1 to 100 and depending on the bin allocates more or less resources to answering it

That’s awesome!

Now have it mark blocks of text on or off, so it can ignore irrelevant, or worse erroneous material — no need to include it in the context window.

Hi, it does not work with llama.cpp right?

Super cool and the results look pretty solid as well! Will give it a try!

i have definitely observed a similar pattern in the Big Label Foundation Models… so, i’m glad to see it in this realm too <3

[dead]

[dead]

[dead]

You didn’t invent this. Models like o3 already do it, that’s why the amount of thinking time varies.