MAROKO133 Update ai: MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI W

📌 MAROKO133 Breaking ai: MIT Researchers Unveil “SEAL”: A New Step Towards Self-Im

The concept of AI self-improvement has been a hot topic in recent research circles, with a flurry of papers emerging and prominent figures like OpenAI CEO Sam Altman weighing in on the future of self-evolving intelligent systems. Now, a new paper from MIT, titled “Self-Adapting Language Models,” introduces SEAL (Self-Adapting LLMs), a novel framework that allows large language models (LLMs) to update their own weights. This development is seen as another significant step towards the realization of truly self-evolving AI.

The research paper, published yesterday, has already ignited considerable discussion, including on Hacker News. SEAL proposes a method where an LLM can generate its own training data through “self-editing” and subsequently update its weights based on new inputs. Crucially, this self-editing process is learned via reinforcement learning, with the reward mechanism tied to the updated model’s downstream performance.

The timing of this paper is particularly notable given the recent surge in interest surrounding AI self-evolution. Earlier this month, several other research efforts garnered attention, including Sakana AI and the University of British Columbia’s “Darwin-Gödel Machine (DGM),” CMU’s “Self-Rewarding Training (SRT),” Shanghai Jiao Tong University’s “MM-UPT” framework for continuous self-improvement in multimodal large models, and the “UI-Genie” self-improvement framework from The Chinese University of Hong Kong in collaboration with vivo.

Adding to the buzz, OpenAI CEO Sam Altman recently shared his vision of a future with self-improving AI and robots in his blog post, “The Gentle Singularity.” He posited that while the initial millions of humanoid robots would need traditional manufacturing, they would then be able to “operate the entire supply chain to build more robots, which can in turn build more chip fabrication facilities, data centers, and so on.” This was quickly followed by a tweet from @VraserX, claiming an OpenAI insider revealed the company was already running recursively self-improving AI internally, a claim that sparked widespread debate about its veracity.

Regardless of the specifics of internal OpenAI developments, the MIT paper on SEAL provides concrete evidence of AI’s progression towards self-evolution.

Understanding SEAL: Self-Adapting Language Models

The core idea behind SEAL is to enable language models to improve themselves when encountering new data by generating their own synthetic data and optimizing their parameters through self-editing. The model’s training objective is to directly generate these self-edits (SEs) using data provided within the model’s context.

The generation of these self-edits is learned through reinforcement learning. The model is rewarded when the generated self-edits, once applied, lead to improved performance on the target task. Therefore, SEAL can be conceptualized as an algorithm with two nested loops: an outer reinforcement learning (RL) loop that optimizes the generation of self-edits, and an inner update loop that uses the generated self-edits to update the model via gradient descent.

This method can be viewed as an instance of meta-learning, where the focus is on how to generate effective self-edits in a meta-learning fashion.

A General Framework

SEAL operates on a single task instance (C,τ), where C is context information relevant to the task, and τ defines the downstream evaluation for assessing the model’s adaptation. For example, in a knowledge integration task, C might be a passage to be integrated into the model’s internal knowledge, and τ a set of questions about that passage.

Given C, the model generates a self-edit SE, which then updates its parameters through supervised fine-tuning: θ′←SFT(θ,SE). Reinforcement learning is used to optimize this self-edit generation: the model performs an action (generates SE), receives a reward r based on LMθ′’s performance on τ, and updates its policy to maximize the expected reward.

The researchers found that traditional online policy methods like GRPO and PPO led to unstable training. They ultimately opted for ReST^EM, a simpler, filtering-based behavioral cloning approach from a DeepMind paper. This method can be viewed as an Expectation-Maximization (EM) process, where the E-step samples candidate outputs from the current model policy, and the M-step reinforces only those samples that yield a positive reward through supervised fine-tuning.

The paper also notes that while the current implementation uses a single model to generate and learn from self-edits, these roles could be separated in a “teacher-student” setup.

Instantiating SEAL in Specific Domains

The MIT team instantiated SEAL in two specific domains: knowledge integration and few-shot learning.

• Knowledge Integration: The goal here is to effectively integrate information from articles into the model’s weights.
• Few-Shot Learning: This involves the model adapting to new tasks with very few examples.

Experimental Results

The experimental results for both few-shot learning and knowledge integration demonstrate the effectiveness of the SEAL framework.

In few-shot learning, using a Llama-3.2-1B-Instruct model, SEAL significantly improved adaptation success rates, achieving 72.5% compared to 20% for models using basic self-edits without RL training, and 0% without adaptation. While still below “Oracle TTT” (an idealized baseline), this indicates substantial progress.

For knowledge integration, using a larger Qwen2.5-7B model to integrate new facts from SQuAD articles, SEAL consistently outperformed baseline methods. Training with synthetically generated data from the base Qwen-2.5-7B model already showed notable improvements, and subsequent reinforcement learning further boosted performance. The accuracy also showed rapid improvement over external RL iterations, often surpassing setups using GPT-4.1 generated data within just two iterations.

Qualitative examples from the paper illustrate how reinforcement learning leads to the generation of more detailed self-edits, resulting in improved performance.

While promising, the researchers also acknowledge some limitations of the SEAL framework, including aspects related to catastrophic forgetting, computational overhead, and context-dependent evaluation. These are discussed in detail in the original paper.

Original Paper: https://arxiv.org/pdf/2506.10943

Project Site: https://jyopari.github.io/posts/seal

Github Repo: https://github.com/Continual-Intelligence/SEAL

The post MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI first appeared on Synced.

🔗 Sumber: syncedreview.com

📌 MAROKO133 Hot ai: China activates massive distributed AI system spanning 1,243 m

China just switched on what may be the world’s largest distributed AI supercomputer, and it spans more than 1,243 miles.

The country has activated a massive, nationwide optical network that links far-flung data centers so efficiently they can work “almost as a single giant computer,” according to Science and Technology Daily.

The system forms a 1,243-mile-wide pool of computing power capable of achieving 98 percent of the efficiency of a single data center, Liu Yunjie, chief director of the project and a member of the Chinese Academy of Engineering, told the state publication.

China’s top computing facilities are scattered across the country. Connected through this new optical backbone, they can now act as a unified machine designed to accelerate AI model training and other compute-heavy research.

Nationwide compute fusion

Liu said the implications of this dedicated, high-speed “data highway” are “revolutionary for scenarios with extremely high real-time demands, such as AI large model training, telemedicine and the industrial internet.”

The backbone is part of the Future Network Test Facility (FNTF), China’s first major national infrastructure project in the information and communication sector. After more than a decade of development, it officially entered operation on December 3.

Researchers say the facility significantly cuts both training time and cost for AI models.

According to Liu, “Training a large model with hundreds of billions of parameters typically requires over 500,000 iterations. On our deterministic network, each iteration takes only about 16 seconds. Without this capability, each iteration would take over 20 seconds longer.”

He also noted that the platform is “ideally positioned to serve the national ‘East Data West Computing’ project,” which shifts data processing to China’s energy-rich western regions.

Decade-long national build

FNTF’s development began in 2013 as part of China’s long-term national science infrastructure roadmap.

The facility now spans 40 cities with more than 34,175 miles of optical transmission lines, enough to wrap around the Earth one and a half times.

Operating around the clock, it supports 128 heterogeneous networks and 4,096 service trials in parallel, making it one of the most extensive testbeds ever deployed.

The project team has created 206 international and domestic standards, secured 221 invention patents, and built what they describe as the world’s first distributed large-scale network operating system.

Liu said the network will eventually be opened to sectors including industrial manufacturing, energy, power, and the low-altitude economy.

Its capabilities were demonstrated at last week’s launch ceremony, when a 72-terabyte dataset from FAST, the world’s largest single-dish radio telescope, was transmitted across 621 miles in just 1.6 hours.

Over the regular internet, the transfer would have taken about 699 days.

The ultra-fast transmission was enabled by a deterministic network channel that guarantees dedicated bandwidth, ultra-low latency, and near-zero packet loss.

When researchers overloaded a parallel 42 Gbps standard channel, speeds collapsed below 1 Gbps, while the 50 Gbps deterministic link stayed at full capacity.

A deterministic network operates “like a precise train timetable,” where every data packet arrives on schedule, state media reported.

Wu Hequan, another member of the Chinese Academy of Engineering involved in evaluating the project, told Science and Technology Daily that the technologies behind FNTF had already supported the development of 5G Advanced and 6G.

“Going forward, both research institutions and enterprises will be able to test various new technologies on this platform,” he said.

🔗 Sumber: interestingengineering.com