Google researchers analyzed Chinese AI models DeepSeek and Alibaba Cloud’s QwQ-32B, finding they exhibit internal multi-agent debates resembling aspects of human collective intelligence.

The study, published on arXiv and not peer-reviewed, suggests that diversity in perspectives, rather than just computational scale, enhances AI reasoning.

The models’ reasoning traces showed interactions such as questioning and perspective-taking, which improved their accuracy when made more conversational.

Alibaba Cloud’s AI unit, Qwen, and DeepSeek are increasingly used in US academia, including at Stanford and Princeton.

The study was led by Google’s “Paradigms of Intelligence” team, with insights from a PhD candidate at the University of Chicago and a Google vice president.

The findings imply that AI models can develop collective reasoning structures through structured interactions among different “voices,” shifting the understanding of AI from solitary problem solvers to collaborative systems.

🔗 Source: South China Morning Post

Recent Google developments