Researchers Claim to Have Discovered a Pathway to Surmount the Data Barrier

Researchers from the Massachusetts Institute of Technology have unveiled a new platform called SEAL, which empowers large language models (LLMs) to generate their own synthetic training data and improve without external assistance.

SEAL operates in two phases. In the first phase, the model learns to create effective «self-editing» models through reinforcement learning. These self-editing models serve as natural language instructions that define new training data and set optimization parameters. During the second phase, the system executes these instructions and updates its weights through machine learning.

A key component of SEAL is the ReST^EM algorithm, which functions as a filter: it retains and amplifies only those modifications that genuinely enhance performance. The algorithm gathers various adjustments, evaluates their effectiveness, and then trains the model exclusively on successful changes. SEAL also leverages low-rank adapters (LoRA), a technique that enables quick and easy updates to the model without the need to retrain it entirely.

The researchers tested SEAL in two scenarios. In the first, they employed Qwen2.5-7B for text comprehension. The model generated logical inferences based on text and then learned from its own outcomes.

SEAL achieved an accuracy of 47%, surpassing the baseline method, which reached 33.5%. The quality of the data it generated even exceeded that of OpenAI’s GPT-4.1, despite the foundational model being significantly smaller.

In the second test, the team utilized Llama 3.2-1B for a reasoning task. Here, the model selected various methods for data processing and training parameters from a predefined toolkit. With the aid of SEAL, the model attained a success rate of 72.5%, compared to only 20% without prior preparation.

Despite these impressive results, the researchers identified several limitations. A primary concern is “catastrophic forgetting,” where the model performs worse on prior tasks as it takes on new challenges. Training also demands substantial resources, with each self-editing evaluation taking between 30 to 45 seconds.

The MIT team views SEAL as a step towards overcoming the so-called “data wall”—the point at which all available human-generated training data is exhausted. Additionally, the researchers caution about the risk of «model collapse,» where model quality diminishes due to overtraining on low-quality AI-generated data. SEAL has the potential to enable continuous learning and the development of autonomous AI systems that consistently adapt to new objectives and information.

If models can learn independently, assimilating new materials—such as research articles—and generating their own explanations and conclusions, they could continue to advance in niche or underexplored subjects. This cycle of self-guided learning may help language models transcend their current capabilities.

The source code for SEAL is available on [GitHub](https://github.com/Continual-Intelligence/SEAL).

You can obtain 100,000 free tokens to complete your initial tasks on [BotHub](https://bothub.chat/?utm_source=contentmarketing&utm_medium=habr&utm_campaign=news&utm_content=RESEARCHERS_SAY_THEY_MAY_HAVE_FOUND_A_WAY_TO_OVERCOME_THE_DATA_WALL) and start working with neural networks right now through [this link](https://bothub.chat/?invitedBy=m_aGCkuyTgqllHCK0dUc7)!

[Source](https://the-decoder.com/researchers-say-they-may-have-found-a-ladder-to-climb-the-data-wall/)