{"id":684,"date":"2025-04-07T20:16:18","date_gmt":"2025-04-07T20:16:18","guid":{"rendered":"https:\/\/www.techrepublic.com\/?p=4301387"},"modified":"2025-04-07T20:16:18","modified_gmt":"2025-04-07T20:16:18","slug":"deepseek-grm-introducing-an-enhanced-ai-reasoning-technique","status":"publish","type":"post","link":"https:\/\/blog.tecnoartesanos.com\/index.php\/2025\/04\/07\/deepseek-grm-introducing-an-enhanced-ai-reasoning-technique\/","title":{"rendered":"DeepSeek-GRM: Introducing an Enhanced AI Reasoning Technique"},"content":{"rendered":"
\"Exevutives
Image: Envato\/DC_Studio<\/figcaption><\/figure>\n

Researchers from AI company DeepSeek and Tsinghua University have introduced a new technique to enhance \u201creasoning\u201d in large language models (LLMs).<\/p>\n

Reasoning capabilities have emerged as a critical benchmark in the race to build top-performing generative AI systems. China and the U.S.<\/a> are actively competing to develop the most powerful and practical models. According to a Stanford University report in April, China\u2019s LLMs are rapidly closing the gap with their U.S. counterparts. In 2024, China produced 15 notable AI models compared to 40 in the U.S., but it leads in patents and academic publications.<\/p>\n

<\/span>What is DeepSeek\u2019s new technique?<\/span><\/h2>\n

DeepSeek researchers published a paper, titled \u201cInference-Time Scaling for Generalist Reward Modeling,\u201d on Cornell University\u2019s arXiv, the archive of scientific papers. Note that papers published on arXiv are not necessarily peer-reviewed.<\/p>\n

In the paper, the researchers detailed a combination of two AI training methods<\/a>: generative reward modeling and self-principled critique tuning.<\/p>\n

\u201cIn this work, we investigate how to improve reward modeling (RM) with more inference compute for general queries, i.e. the inference-time scalability of generalist RM, and further, how to improve the effectiveness of performance-compute scaling with proper learning methods,\u201d the researchers wrote.<\/p>\n