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

📌 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 Breaking ai: It Seems Almost Statistically Impossible That This Polyma

Earlier this year, an anonymous bettor on Polymarket perfectly predicted the US invasion of Venezuela mere hours before over 150 US aircraft rocked the country’s capital of Caracas, netting them over $400,000.

The incident reignited a heated debate over insider trading on prediction market platforms like Polymarket and Kalshi. While the act is strictly forbidden on Wall Street, prediction markets are currently operating in a regulatory vacuum, allowing those who enjoy insider status to score big — while everyone else is left to pick up the bill.

And the evidence that prediction markets are rife with insider traders continues to grow. As one eagle-eyed Reddit user noticed, an anonymous day-old Polymarket account correctly guessed 17 out of around 20 bets about Sunday’s Super Bowl half-time show.

Statistically speaking, that’s an exceedingly unlikely success rate, strongly suggesting the account had some kind of insider knowledge of what would happen during Puerto Rican superstar Bad Bunny’s performance.

The account correctly predicted that popstars Lady Gaga, Cardi B, and Ricky Martin would perform at the show. The account also correctly predicted that rappers Travis Scott and Drake, as well as singer Post Malone, would not perform.

The anonymous user placed bets starting February 6, two days before the event took place. All told, they made about $17,000 in profit.

“If you bet, you’re a rube for these people,” one Reddit user commented. “Literally spending money to give it over to insiders and cheats.”

Polymarket has yet to publicly comment on the matter and didn’t respond to the Wall Street Journal‘s request for comment.

Ironically, gambling was a major focus during Sunday’s sporting event. Companies spent untold sums to air commercials for betting services like DraftKings.

Trading volumes on prediction markets hit record highs during the Super Bowl, with Kalshi reporting over half a billion dollars in trading tied to the final outcome of the game. Polymarket’s equivalent bet saw trading volumes hit over $55 million.

Given ongoing regulatory uncertainty, it remains to be seen whether the two platforms will be able to meaningfully address insider trading, a subject that demonstrably continues to be a major problem.

While a knowledgeable few, including professional gamblers, get away with major profits, plenty of other users are incurring significant losses. As Business Insider points out, for each “winner” on the platform, there’s a “loser” as well, since users are betting against each other — not against a house, like in a traditional casino.

The latter will need to remain confident that the setup isn’t rigged against them for the ruse to work.

“Those other people are going to, on average, make losses if they know less about the subject matter than the experts,” professor of economics and prediction markets expert Eric Zitzewitz told BI. “So you need them to be willing to trade despite that.”

“You need them to be overconfident about how much they know, or you need them to be participating for some other reason,” he added.

Prediction market regulations could take years before they materialize — if ever. The White House has made it clear it supports the industry, with president Donald Trump’s Trump Media and Technology Group announcing in October that it would enter the prediction markets business.

In the meantime, lawmakers are desperately warning that gambling on the platforms comes with some inherent risks.

“New Yorkers need to know the significant risks with unregulated prediction markets,” New York Attorney General Letitia James warned in a statement six days ahead of the Super Bowl. “It’s crystal clear: so-called prediction markets do not have the same consumer protections as regulated platforms. I urge all New Yorkers to be cautious of these platforms to protect their money.”

More on prediction markets: Professional Gamblers Move Into Prediction Markets to Bleed You Dry

The post It Seems Almost Statistically Impossible That This Polymarket Bettor Didn’t Profit Off Inside Knowledge About the Super Bowl Half Time Show appeared first on Futurism.

🔗 Sumber: futurism.com