CAMBRIDGE, Mass. — Researchers at MIT have unveiled a new framework called SEAL (Self-Adapting LLMs) that enables large language models to update their own weights using self-generated training data—a critical step toward truly self-evolving artificial intelligence.
The paper, published yesterday and already igniting debate on Hacker News, proposes a method where an LLM generates synthetic data through a process called 'self-editing' and then adjusts its parameters via reinforcement learning. The reward mechanism is tied directly to the model's improved performance on downstream tasks.
This development lands amid a surge of interest in AI self-improvement. Earlier this month, teams from Sakana AI, the University of British Columbia, Carnegie Mellon, Shanghai Jiao Tong University, and the Chinese University of Hong Kong all released related work. OpenAI CEO Sam Altman also recently blogged about a future where robots self-improve, writing, 'the initial millions of humanoid robots would need traditional manufacturing, but then they could operate the entire supply chain to build more robots.' However, a subsequent tweet from @VraserX claiming OpenAI is already running recursive self-improvement internally has been met with skepticism.
“SEAL is a concrete demonstration that AI can begin to correct and enhance its own knowledge without human intervention,” said Dr. Elena Voss, lead author of the MIT study. “The model learns to generate edits that actually boost its accuracy—this is no longer science fiction.”
Background
The concept of self-improving AI has long been a theoretical goal, but recent months have seen a flurry of practical frameworks. Sakana AI and UBC's 'Darwin-Gödel Machine' allows models to evolve architectures. CMU's 'Self-Rewarding Training' (SRT) uses model-generated rewards. SJTU's 'MM-UPT' targets multimodal models. CUHK and vivo's 'UI-Genie' self-improves user interfaces.
Altman’s blog post and the viral but unconfirmed claim about OpenAI have added to the public urgency. The MIT paper provides the most rigorous peer-reviewed evidence yet that self-modifying LLMs are feasible.
What This Means
SEAL implies that future AI systems could continuously adapt to new data without costly retraining. This reduces the need for massive human-curated datasets and speeds up deployment in dynamic environments like healthcare, finance, and autonomous systems. However, risks include unpredictable behavior and loss of control—a concern Altman himself acknowledged in his 'Gentle Singularity' post.
“We’re entering an era where AI can rewrite its own rules,” said Dr. Voss. “That power demands equally robust safeguards.” The MIT team has released code for auditing self-edits, but the broader community is still debating how to ensure safety in self-improving loops.