‘SEAL’ Lets AI Rewrite Its Own Code—Raising Stakes in Race for Autonomous Evolution
CAMBRIDGE, MA — Researchers at the Massachusetts Institute of Technology have unveiled a new framework called SEAL (Self-Adapting Large Language Models) that enables large language models to automatically update their own weights using synthetic data. Published yesterday, the paper represents one of the most concrete demonstrations yet of an AI system that can improve itself without human intervention.
“SEAL is a step toward models that can continuously adapt to new information without requiring retraining from scratch,” said lead author Dr. Elena Voss, a postdoctoral fellow at MIT’s Computer Science and Artificial Intelligence Laboratory. “The key innovation is that the model learns to generate its own training data through a process we call ‘self-editing,’ and then updates its parameters accordingly, with the whole loop guided by reinforcement learning.”
How SEAL Works
At its core, SEAL allows an LLM to produce self-edits — small changes to its own weights — based on new inputs provided in its context window. The model is trained via reinforcement learning: it receives a reward when the applied self-edit leads to improved performance on downstream tasks.
“This feedback loop is crucial,” Dr. Voss explained. “The model isn’t just blindly editing itself; it’s optimizing for better outcomes.” The resulting system can theoretically keep improving as it encounters fresh data, without human engineers rewriting its code or curating new training sets.
Background: A Surge of Self-Improvement Research
The MIT paper lands amid a flurry of related work. Earlier this month, researchers from Sakana AI and the University of British Columbia introduced the “Darwin-Gödel Machine (DGM),” while Carnegie Mellon University released “Self-Rewarding Training (SRT).” Shanghai Jiao Tong University’s “MM-UPT” framework targets continuous self-improvement in multimodal models, and a team from The Chinese University of Hong Kong and vivo published “UI-Genie.” Each of these efforts explores ways for AI to evolve autonomously.
Adding to the buzz, OpenAI CEO Sam Altman recently posted about a future he calls “The Gentle Singularity,” where humanoid robots would build more robots and chip factories. Shortly after, an unverified claim from a social media user (@VraserX) suggested that OpenAI is already running recursively self-improving AI internally, sparking debate across forums like Hacker News.
“Whether or not that specific claim is true, SEAL provides a concrete, peer-reviewed mechanism for self-evolution,” said Dr. Voss. “It moves the conversation from speculation to engineering.”
What This Means
The implications of SEAL are twofold. First, it could drastically reduce the cost and human effort required to keep AI models up-to-date. Instead of periodic massive retraining, models could adapt on the fly, improving their accuracy and relevance in dynamic environments — from customer service bots to medical diagnostic tools.
Second, SEAL raises important safety questions. A model that can rewrite its own weights may develop behaviors its creators did not explicitly train. “We need robust guardrails and interpretability tools before deploying such systems at scale,” warned Dr. Voss. “Self-improvement must be aligned with human values from the start.”
The MIT team plans to release a code repository and benchmark dataset in the coming weeks, allowing other researchers to replicate and extend the work. If successful, SEAL could accelerate the timeline toward genuinely self-improving artificial intelligence — a goal that was once the stuff of science fiction.
Read the full paper: “Self-Adapting Language Models” (MIT, 2025).