A Chinese quantitative trading fund has submitted a research paper to one of the world’s top artificial intelligence (AI) conferences, detailing a new training technique that it said could outperform mainstream methods employed by leading AI research labs, in a move that mirrors the path taken by DeepSeek.

Shanghai Goku Technologies, established in 2015, submitted the paper to the Conference on Neural Information Processing Systems – an annual gathering of top scientists in machine learning and AI that is often referred to as the “AI Olympics”.

In its paper, Goku laid out the limits of popular AI training methods – including supervised fine-tuning (SFT) and reinforcement learning (RL) – and proposed a so-called step-wise adaptive hybrid training framework called SASR, which it said was inspired by the way humans develop reasoning capabilities.

SFT and RL are key techniques used by companies such as Microsoft-backed OpenAI and DeepSeek to train their AI models. DeepSeek previously highlighted the significance of SFT and RL in enhancing the performance of its V3 model, which made waves in the global technology community upon release in December.

“Experimental results demonstrate that SASR outperforms SFT, RL and static hybrid training methods,” the Goku team wrote in its paper, which was co-authored with researchers from Shanghai Jiao Tong University and Goku’s newly established AI subsidiary, Shanghai AllMind Artificial Intelligence Technology.

Goku, which operates under the slogan “logic and truth are the only principles we obey”, did not immediately respond to a request for comment on Thursday.