Teaching large language models how to absorb new knowledge – MIT News

Advancements in Self-Adapting AI and Implications for Sustainable Development Goals

Introduction

A recent study by MIT researchers introduces a novel framework enabling Large Language Models (LLMs) to permanently internalize new information, overcoming their traditionally static nature post-deployment. This development, which allows an AI to learn and adapt continuously, holds significant potential for accelerating progress toward several United Nations Sustainable Development Goals (SDGs). By creating AI systems that can evolve in response to new data and changing environments, this research paves the way for more effective and dynamic tools to address global challenges in education, innovation, and economic growth.

Methodology: The SEAL Framework

The new approach, termed SEAL (Self-Adapting LLMs), leverages a model’s inherent in-context learning capabilities to facilitate permanent updates to its core knowledge base. The process emulates a human student’s study habits.

1. Data Generation: The LLM autonomously generates its own “study sheets” by rewriting and summarizing new input information into multiple synthetic data points. Each piece of synthetic data represents a potential self-edit.
2. Trial-and-Error Learning: Using reinforcement learning, the model assesses each self-edit to determine which one yields the greatest performance improvement on a given task, such as question answering. The model is rewarded for identifying the most effective update.
3. Knowledge Internalization: The LLM permanently updates its internal weights based on the most effective self-edit, thereby memorizing and internalizing the new information.

Crucially, the SEAL framework also allows the model to control its own learning process, including selecting the data to learn from and configuring the optimization parameters. This grants the model the ability to determine the most efficient way to digest and integrate new knowledge.

Performance and Key Findings

The SEAL method demonstrated superior performance compared to baseline models across a range of tasks. Key results include:

• A nearly 15% improvement in model accuracy on question-answering tasks.
• An increase in success rate of over 50% on certain skill-learning tasks.
• The enablement of a smaller model to outperform much larger, conventionally static LLMs.

Despite these successes, researchers identified “catastrophic forgetting”—the gradual decline in performance on earlier tasks as new information is learned—as a limitation requiring further investigation.

Alignment with Sustainable Development Goals

The development of self-adapting AI directly supports the achievement of key Sustainable Development Goals by providing a foundation for more intelligent, responsive, and efficient systems.

• SDG 4: Quality Education

  The AI’s ability to continuously learn serves as a powerful model for next-generation educational technologies. Self-adapting systems can power personalized learning platforms that remain current with the latest information, promoting lifelong learning opportunities and providing equitable access to quality educational content that evolves with scientific and cultural knowledge.

• SDG 9: Industry, Innovation, and Infrastructure

  This research is a fundamental innovation that enhances digital infrastructure. Adaptable AI that can integrate new data without complete retraining can accelerate scientific discovery and foster sustainable industrial processes. This contributes to building resilient infrastructure and promoting inclusive and sustainable industrialization by creating tools that continuously improve their capabilities.

• SDG 8: Decent Work and Economic Growth

  By improving the accuracy and adaptability of AI, this technology can boost productivity and drive sustained, inclusive economic growth. AI agents that can learn new skills on the fly can support human workers, automate complex tasks more effectively, and contribute to the creation of higher-value jobs. This aligns with the goal of achieving full and productive employment and decent work for all.

• SDG 17: Partnerships for the Goals

  The collaborative nature of this research, supported by various institutions, exemplifies the spirit of partnership required to achieve the SDGs. Furthermore, the future research direction of creating multi-agent systems where LLMs train each other offers a technological parallel to global partnerships, demonstrating how collaborative networks can be leveraged to solve complex problems and advance shared goals.

Sustainable Development Goals (SDGs) Addressed in the Article

SDG 9: Industry, Innovation, and Infrastructure

• The article’s core subject is a significant technological innovation in artificial intelligence developed by researchers at MIT. This directly relates to fostering innovation and upgrading technological capabilities. The text highlights a “new approach developed by MIT researchers” that could “help artificial intelligence agents consistently adapt to new tasks” and ultimately “help advance science.”

SDG 4: Quality Education

• The research is situated within an educational institution (MIT) and involves “an MIT graduate student” and “an MIT undergraduate” as lead authors. The entire process of the AI learning is explained using an educational analogy: “a professor lectures while students diligently write down notes,” and the AI “generates its own study sheets.” This underscores the role of higher education in driving innovation and developing advanced technical skills.

SDG 17: Partnerships for the Goals

• The article explicitly mentions that the research is a collaborative effort supported by multiple stakeholders. The final paragraph states, “This work is supported, in part, by the U.S. Army Research Office, the U.S. Air Force AI Accelerator, the Stevens Fund for MIT UROP, and the MIT-IBM Watson AI Lab.” This demonstrates a public-private partnership model to advance science and technology.

Specific SDG Targets Identified

Targets under SDG 9: Industry, Innovation, and Infrastructure

1. Target 9.5: Enhance scientific research, upgrade the technological capabilities of industrial sectors in all countries… encouraging innovation and substantially increasing the number of research and development workers.
   • The article is a direct manifestation of this target. It describes fundamental scientific research (“The research will be presented at the Conference on Neural Information Processing Systems”) that upgrades the technological capabilities of AI. The project team, consisting of professors, graduate students, and undergraduates, represents the research and development workers central to this target.

Targets under SDG 4: Quality Education

1. Target 4.4: By 2030, substantially increase the number of youth and adults who have relevant skills, including technical and vocational skills, for employment, decent jobs and entrepreneurship.
   • The research team, which includes an undergraduate and graduate students, is actively acquiring and developing highly relevant, cutting-edge technical skills in artificial intelligence and machine learning. Their work on “self-adapting LLMs” is at the forefront of ICT skills development.

Targets under SDG 17: Partnerships for the Goals

1. Target 17.17: Encourage and promote effective public, public-private and civil society partnerships, building on the experience and resourcing strategies of partnerships.
   • The funding and support structure mentioned in the article is a clear example of this target in action. The collaboration between a university (MIT), government entities (U.S. Army, U.S. Air Force), and a private corporation (IBM via the MIT-IBM Watson AI Lab) showcases a multi-stakeholder partnership mobilizing financial resources and expertise for technological innovation.

Implied Indicators for Measuring Progress

Indicators for SDG 9

• Implied Indicator 9.5.2 (Researchers per million inhabitants): The article names a team of researchers, including “an MIT graduate student,” “an MIT undergraduate,” and “senior authors Yoon Kim… and Pulkit Agrawal.” This highlights the human capital—the researchers themselves—invested in driving innovation, which is the basis of this indicator.
• Implied Indicator 9.5.1 (Research and development expenditure as a proportion of GDP): The article’s acknowledgment of financial support from the “U.S. Army Research Office, the U.S. Air Force AI Accelerator, and the MIT-IBM Watson AI Lab” directly implies significant research and development expenditure, which this indicator aims to measure.

Indicators for SDG 4

• Implied Indicator 4.4.1 (Proportion of youth and adults with information and communications technology (ICT) skills): The student researchers involved in creating this advanced AI system are prime examples of youth and adults possessing exceptionally high-level ICT skills. The technology itself represents a new frontier in ICT, and those who develop it are the measure of this skilled population.

Indicators for SDG 17

• Implied Indicator 17.17.1 (Amount of United States dollars committed to public-private and civil society partnerships): While no specific dollar amount is given, the mention of multiple, well-funded organizations (U.S. military research offices, IBM) supporting the project implies a significant financial commitment to this public-private partnership for research and development. The existence and naming of the partnership itself serves as a qualitative measure of progress toward this target.

Summary of Findings

Source: news.mit.edu