AI models are stuck “too thoughtful”. Nvidia and Google have a solution.

Large language models – they are like us. Or at least, they are trained to answer like us. And now they even display some of the most annoying features that accompany reasoning capacities and “reflection”.

Reasoning models like O1 of Openai or R1 from Deepseek have been trained to question their logic and check their own answers. But if they do it too long, the quality of the answers they generate begins to deteriorate.

“The longer he thinks, the more likely it is that the answer is deceived because it is hiding,” Jared Quincy Davis, founder and CEO of Foundry, told Business Insider. Relatively, right?

“It’s as if a student is going to examine and spends three hours on the first question. It’s too thoughtful – it’s stuck in a loop,” continued Davis.

Davis, as well as researchers from Nvidia, Google, IBM, MIT, Stanford, Databricks, etc., launched an open source frame on Tuesday. This is called EMPER and he could presume the next phase of large -language models.

Overwhelming and decreasing reversal

The concept of “too thoughtful” may seem to contradict another great rupture of the improvement of the model: the scaling of inference. Barely a few months ago, the models that took a little more time to find a more considered response were presented by AI luminaires like Jensen Huang as the future of model improvement.

The models of reasoning and the time scaling are always a huge step, said Davis, but future developers will probably think of using them, and all models, differently.

Davis and the EMPER team formalize a structure around a concept with which he and other AI researchers have been playing for months.

Nine months ago – an eon in the world of automatic learning – Davis described its hacking of asking, called “call”, Chatgpt 4 the same question on several occasions and take the best of the answers.

Now, the researchers from Ember take this method and overcome it, imagining compound systems in which each question or task would call a patchwork of models, all for different quantities of reflection time, depending on what is optimal for each model and each question.

“Our system is a framework to build these networks of networks where you want, for example, to deal with many calls in a broader system that has its own properties. It is therefore like a new discipline which, I think, jumped from research to practice very quickly,” said Davis.

Related stories

Business Insider tells the innovative stories you want to know

Business Insider tells the innovative stories you want to know

In the future, the model will choose you

When humans reflect, therapists could tell us to break up problems into small pieces and solve them at a time. EMBER shares a part of this theory, but the similarity ends fairly quickly.

Currently, when you log into Perplexity or Chatgpt, you choose your model with a drop -down menu or a rocking switch. Davis thinks that this will not be the case much longer, because IA companies are looking for better results with these more complex routing strategies of questions through different models with different numbers and lengths of calls.

“You can imagine, instead of being a million calls, it could be a billion of calls or quadrillion calls. You must sort the calls,” said Davis. “You have to choose models for everyone call. Should each call be GPT 4? Or should some calls be GPT 3? Should some calls be anthropogenic or Gemini, and others call Deepseek? What should be the prompts for each call? “”

It is to think in more dimensions than the binary questions and answers that we have known. And this will be particularly important while we move to the AIA agents where the models perform tasks without human intervention.

Davis compared these AI systems compounds with chemical genius.

“It’s a new science,” he said.

Do you have a tip or an overview to share? Contact Emma at ecosgro@businessinsider.com or use the secure messaging application signal: 443-333-9088