DeepSeek reveals a “revolutionary” technology to operate its upcoming R3 model for inference [Arabic]

Author: Asharq News | Source: Asharq News | Read the full article in Arabic

DeepSeek, a Chinese startup focused on artificial intelligence, has unveiled an innovative approach aimed at enhancing the reasoning capabilities of large language models (LLMs). This announcement comes as anticipation builds for the release of the company's next-generation models. Collaborating with researchers from Tsinghua University, DeepSeek introduced two new methodologies designed to improve the accuracy and speed of responses to general questions.

The first methodology, known as Generative Reward Modeling (GRM), teaches AI models how to align their answers with human preferences. Instead of requiring human evaluation for every response, GRM enables a secondary model to automatically assess answers and provide feedback based on their quality. This process is akin to a game where the AI earns points for good answers and loses points for incorrect ones, ultimately refining its ability to respond effectively.

Researchers have reported that the new DeepSeek-GRM models have outperformed existing methods, achieving competitive results compared to other high-performing reward models. The company plans to release open-source versions of the GRM models, although a specific timeline has not yet been established. This development comes amid growing speculation about DeepSeek's future steps, especially following the success of its previous models, which have garnered significant attention in the tech community.

[Read More (translated)]