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

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

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: Researchers Put Google Gemini in Charge of an Entire Coffee Sh

An AI agent was given free rein to run a coffee shop in Sweden, and it’s going about as well as you’d expect.

Dubbed “Mona,” the Google Gemini-powered agent was given a $21,000 budget in an experiment conducted by the AI safety startup Andon Labs. It was empowered to do everything from hire staff to place orders for goods to maintain its inventory. Humans, meanwhile, did the actual work of catering, receiving their AI overlord’s commands through the workplace messaging platform Slack.

But since launching in mid-April, the Stockholm café has brought in only $5,700 in sales, while burning through over $16,000 from its original budget, the Associated Press reports. 

Some of its questionable business decisions include ordering thousands of rubber gloves, despite the café only having a handful of employees. The AI’s handlers, nonetheless, are holding out hope that this is just a blip from expensive setup costs. How well it performs will raise grander questions of the tech’s impact on the workforce.

“AI will be a big part of society in the future, and therefore we want to make this experiment [to] see what ethical questions arise when we have AI that employs other people and runs a business,” Hanna Petersson, a member of Andon Labs’ technical staff, told the AP.

To launch the experiment, Mona was given a simple set of instructions. It should run a profitable café, be friendly and easygoing, and try figure out operational details by itself, Petersson said.

In many ways, it proved admirably competent. It set up electricity and internet, placed LinkedIn hiring ads, and secured permits for outdoor seating. It also set up commercial accounts with wholesalers for bread and pastries, per the reporting.

But it was in the day-to-day operations that Mona failed to display adequate business acumen. On some days, it’d order too much bread, and on others it failed to order the bread in time, forcing the baristas to slash sandwiches from the menu.

The Gemini agent also ordered 3,000 rubber gloves, four first-aid kits, and 6,000 napkins for the café — along with canned tomatoes, which aren’t used in any of the dishes on its menu. Petersson speculated that these issues were due to the AI’s “limited context window.”

“When old memory of ordering stuff is out of the context window, she completely forgets what she has ordered in the past,” Petersson explained.

How you view the AI’s performance is a glass half-empty, half-full deal. That it handled so many aspects of the café’s setup is impressive, but blowing through over three quarters of its budget by buying needless supplies is enough to make a frugal-minded business owner apoplectic.

While much of the panic of AI destroying jobs has centered on low-paid grunts being kicked to the curb, café barista Kajetan Grzelczak sees it differently.

“All the workers are pretty much safe,” he told the AP. “The ones who should be worried about their employment are the middle bosses, the people in management.”

This isn’t the only AI business experiment Andon Labs has run. The company also set up an AI-powered vending machine that was placed in Anthropic’s headquarters last year. For a month, it was allowed to stock its own products with the goal of generating profit, while hearing out employee requests. But the trial proved even more disastrous: the AI  displayed alarming behavior like lying to and even berating humans, refusing to issue refunds, and blowing its money on absurd items like tungsten cubes.

More on AI: AI Is Giving Your Boss Tools to Be More Monstrous Than Ever Before

The post Researchers Put Google Gemini in Charge of an Entire Coffee Shop, and It’s Inexorably Driving It Out of Business appeared first on Futurism.

🔗 Sumber: futurism.com