SEOUL, South Korea – VIDRAFT, an AI startup based at Seoul AI Hub, has announced the global release of MARL, or Model-Agnostic Runtime Middleware for LLMs. This innovative reasoning middleware aims to significantly reduce hallucinations and enhance self-correction in large language models (LLMs). MARL is now accessible on platforms including Hugging Face, GitHub, PyPI, and ClawHub, the skill marketplace associated with the AI agent platform OpenClaw.
MARL serves as a runtime layer that enhances the reasoning capabilities of language models without necessitating fine-tuning or retraining. Developers and enterprises can implement MARL with minimal code alterations across a variety of models that align with the OpenAI API format. This includes widely recognized models such as GPT, Claude, Gemini, DeepSeek, Grok, and Llama.
According to VIDRAFT, many LLMs frequently deliver incorrect responses with high confidence, often lacking robust mechanisms for error detection and correction. MARL addresses this challenge by restructuring a model’s response into a multi-stage reasoning process. This method first involves planning an approach, followed by drafting an answer, conducting independent verification of that draft, and ultimately generating a revised final response. The outcome is a more deliberate and consistent output compared to the traditional single-pass generation process.
The company’s internal evaluation framework, known as FINAL Bench, highlighted a noticeable gap among current frontier models in recognizing and rectifying potential errors. VIDRAFT reported that MARL demonstrated substantial improvements in more complex tasks, with a significant portion of the progress attributed to its self-correction phase.
In addition to general reasoning enhancements, MARL incorporates specialized reasoning engines tailored for specific domains such as drug discovery, legal analysis, and creative work. This feature enables organizations to increase response reliability without altering model weights or committing to a single model vendor.
Moreover, MARL has been listed on ClawHub, which serves as a marketplace for OpenClaw. VIDRAFT emphasized that while agent platforms are intended to execute tasks, MARL enhances the reasoning quality underpinning those tasks by adding a structured layer of reflection and verification prior to delivering a final answer.
“We started with a simple observation: even the most advanced AI models still struggle to reliably assess the limits of their own answers,” remarked Min-Sik Kim, CEO of VIDRAFT. “MARL does not replace the model. It changes how the model thinks at runtime. We believe trustworthy AI begins with systems that can question, review, and improve their own outputs before presenting them to users.”
VIDRAFT is in the process of preparing an enterprise edition of MARL, which is expected to launch in the first half of 2026. The company plans to submit its validation results for academic publication and has already conducted proof-of-concept engagements in the U.S. market, with ongoing efforts in localization.
About VIDRAFT
Founded in 2024, VIDRAFT is an AI startup focused on advancing toward True AGI by 2030. The company developed FINAL Bench, an evaluation benchmark for AI metacognition, and has achieved a series of results across global AI platforms and leaderboards.
Media Contact
Company Name: VIDRAFT
Contact Person: CHUNG HOON LEE
Email: Send Email
Country: South Korea
Website: https://vidraft.net/
Zendaya Addresses AI Wedding Photo Controversy While Promoting New Film With Pattinson
U.S. Appeals Court Allows Perplexity AI’s Comet to Operate on Amazon Amid Legal Battle
Microsoft Considers Legal Action Against Amazon and OpenAI Over $50B AI Deal
Microsoft Mulls Legal Action Over $50B Amazon-OpenAI Cloud Deal Amid Rising AI Competition
Mistral Launches Forge for Customized AI Models Amid Growing Data Sovereignty Needs
