Big modern language Models (LLMs) may write beautiful sonnets and elegant codes, but they don't even have the basic ability to learn from experience.
Researchers at the Massachusetts Institute of Technology (MIT) have devised ways that LLM continues to improve by tweaking their own parameters in response to useful new information.
This task is a step towards building an artificial intelligence model that is constantly learning. This is important when the field has long-standing goals and when machines are more faithful to mimic human intelligence. In the meantime, we can provide chatbots and other AI tools that can better incorporate new information, including user interests and preferences.
The MIT scheme, called the Self-Adaptive Language Model (SEAL), involves generating LLMs' own synthetic training data and learning update procedures based on the inputs received.
“The first idea was to find out if it was a token or not. [units of text fed to LLMs and generated by them] It could lead to strong model updates,” says Jyothish Pari, a doctoral student at MIT, involved in the development of SEAL. Paris says it's about checking if you can train using the output of the model.
Adam Zweiger, a researcher at the MIT department involved in building SEAL, adds that new models can “infer” into better solutions by implementing more complex inferences, but the model itself will not benefit from this inference in the long run.
In contrast, seals generate new insights and fold them into their own weights or parameters. For example, in light of statements about the challenges facing the Apollo Space Program, the model generated new sentences that sought to explain the meaning of the statement. Researchers compared this to how human students write and review notes to help them learn.
The system then updated the model with this data to test how well the new model answered a set of questions. And finally, this provides reinforcement learning signals that help guide the model towards updates that will help improve overall competence and continue learning.
The researchers tested the approach with small and medium sized versions of two open source models: Meta's Lama and Alibaba's Qwen. They say this approach should work with a much larger frontier model as well.
Researchers tested a textual seal approach and a benchmark called ARC to measure the ability of AI models to solve abstract inference problems. In both cases, we found that the seals allow the model to continue learning beyond initial training.
Pulkit Agrawal, a professor at MIT who directed the film, says the SEAL project touches on key themes in AI. He says it can be used to help make AI models more personalized. “LLM is strong, but we don't want to stop their knowledge,” he says.
Seals are not a way for AI to improve indefinitely. For one thing, as Agrawal points out, the tested LLMS suffers from what is known as “catastrophic forgetting.” Ingesting new information has the troublesome effect of simply disappearing old knowledge. This may indicate the fundamental difference between artificial neural networks and biological networks. Pari and Zweigler also note that Seal is computationally intensive and the best way to schedule new learning periods is not yet clear. One fun idea is that, like humans, LLM can probably experience a period of “sleep” where new information is integrated.
Still, with all its limitations, SEAL is an exciting new path for further AI research. It may be something that finds a way to a future frontier AI model.
What do you think about AI that allows you to continue learning? Please email hello@wired.com.
