Large language models (LLMs) like ChatGPT
Large language models (LLMs) like ChatGPT can write an essay or plan a menu almost instantly. But until recently, it was also easy to stump them. The models, which rely on language patterns to respond to users’ queries, often failed at math problems and were not good at complex reasoning. Suddenly, however, they’ve gotten a lot better at these things.
A new generation of LLMs known as reasoning models
A new generation of LLMs known as reasoning models are being trained to solve complex problems. Like humans, they need some time to think through problems like these — and remarkably, scientists at MIT’s McGovern Institute for Brain Research have found that the kinds of problems that require the most processing from reasoning models are the very same problems that people need take their time with. In other words, they report today in the journal PNAS, the “cost of thinking” for a reasoning model is similar to the cost of thinking for a human.
The researchers, who were led by Evelina Fedorenko, an associate professor of brain and cognitive sciences and an investigator at the McGovern Institute, conclude that in at least one important way, reasoning models have a human-like approach to thinking. That, they note, is not by design. “People who build these models don’t care if they do it like humans. They just want a system that will robustly perform under all sorts of conditions and produce correct responses,” Fedorenko says. “The fact that there’s some convergence is really quite striking.”
Reasoning models
Like many forms of artificial intelligence, the new reasoning models are artificial neural networks: computational tools that learn how to process information when they are given data and a problem to solve. Artificial neural networks have been very successful at many of the tasks that the brain’s own neural networks do well — and in some cases, neuroscientists have discovered that those that perform best do share certain aspects of information processing in the brain. Still, some scientists argued that artificial intelligence was not ready to take on more sophisticated aspects of human intelligence.
“Up until recently, I was among the people saying, ‘These models are really good at things like perception and language, but it’s still going to be a long ways off until we have neural network models that can do reasoning,” Fedorenko says. “Then these large reasoning models emerged and they seem to do much better at a lot of these thinking tasks, like solving math problems and writing pieces of computer code.”
Andrea Gregor de Varda, a K. Lisa Yang ICoN Center Fellow and a postdoc in Fedorenko’s lab, explains that reasoning models work out problems step by step. “At some point, people realized that models needed to have more space to perform the actual computations that are needed to solve complex problems,” he says. “The performance started becoming way, way stronger if you let the models break down the problems into parts.”
To encourage models to work through complex problems in steps that lead to correct solutions, engineers can use reinforcement learning. During their training, the models are rewarded for correct answers and penalized for wrong ones. “The models explore the problem space themselves,” de Varda says. “The actions that lead to positive rewards are reinforced, so that they produce correct solutions more often.”
Models trained in this way are much more likely than their predecessors to arrive at the same answers a human would when they are given a reasoning task. Their stepwise problem-solving does mean reasoning models can take a bit longer to find an answer than the LLMs that came before — but since they’re getting right answers where the previous models would have failed, their responses are worth the wait.
The models’ need to take some time to work through complex problems already hints at a parallel to human thinking: if you demand that a person solve a hard problem instantaneously, they’d probably fail, too. De Varda wanted to examine this relationship more systematically. So he gave reasoning models and human volunteers the same set of problems, and tracked not just whether they got the answers right, but also how much time or effort it took them to get there.
Time versus tokens
This meant measuring how long it took people to respond to each question, down to the millisecond. For the models, Varda used a different metric. It didn’t make sense to measure processing time, since this is more dependent on computer hardware than the effort the model puts into solving a problem. So instead, he tracked tokens, which are part of a model’s internal chain of thought. “They produce tokens that are not meant for the user to see and work on, but just to have some track of the internal computation that they’re doing,” de Varda explains. “It’s as if they were talking to themselves.”
Both humans and reasoning modelsKindly read our copyright disclaimer here: https://cere-sync.com/dmca-copyrights-disclaimer/