SEOUL, South Korea - VIDRAFT, an AI startup based at Seoul AI Hub, today announced the global release of MARL, short for Model-Agnostic Runtime Middleware for LLMs, a reasoning middleware designed to reduce hallucinations and strengthen self-correction in large language models. MARL is now available on Hugging Face, GitHub, PyPI, and ClawHub, the skill marketplace of the AI agent platform OpenClaw.
MARL is a runtime layer that improves how language models reason without requiring fine-tuning or retraining. With minimal code changes, developers and enterprises can apply MARL across a wide range of models that support the OpenAI API format, including GPT models, Claude, Gemini, DeepSeek, Grok, and Llama
According to VIDRAFT, many LLMs still produce incorrect answers with high confidence while lacking a reliable mechanism to catch and correct their own errors. MARL addresses that limitation by restructuring a single model response into a multi-stage reasoning process. It first plans an approach, then drafts an answer, subjects that draft to an independent verification step, and finally generates a revised final response. The result is a more deliberate and consistent output than a standard single-pass generation flow.
The company said its internal evaluation framework, FINAL Bench, found a significant gap in today’s frontier models between recognizing the possibility of error and actually correcting it. VIDRAFT said MARL showed the strongest gains on more difficult tasks, with much of the improvement coming from the self-correction stage.
In addition to general reasoning enhancement, MARL includes specialized reasoning engines tailored for domains such as drug discovery, legal analysis, and creative work. VIDRAFT said this allows organizations to improve response reliability without changing model weights or locking themselves into a single model vendor.
MARL has also been listed on ClawHub, the marketplace for OpenClaw. VIDRAFT said that if agent platforms are built to execute tasks, MARL is designed to strengthen the reasoning quality behind those actions by adding a structured layer of reflection and verification before a final answer is produced.
“We started with a simple observation: even the most advanced AI models still struggle to reliably assess the limits of their own answers,” said 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 preparing to launch an enterprise edition of MARL in the first half of 2026 and plans to submit its validation results for academic publication. The company said it has also completed proof-of-concept engagements for the U.S. market and is currently working on 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 recorded 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/
