Machine Reasoning and Deep Thinking: Insights from Ant Financial’s NLP Lead Wu Wei

After the release of DeepSeek R1, the AI community witnessed a surge of interest in long‑thinking chains, which demonstrate that large models can perform reasoning that appears more human‑like and even surpass human capabilities in certain tasks.

Wu Wei, head of natural‑language processing at Ant Financial’s Technology Research Institute, reflects on the historical progression from early NLP systems that merely "processed" language to modern models that "understand" and generate text, citing milestones such as BERT (2018), Meena (2020), and ChatGPT (2022).

He defines machine reasoning as the algebraic manipulation of existing knowledge to solve new problems, echoing Leon Bottou’s 2011 definition and Marvin Minsky’s concept of intelligent agents that combine to form higher‑level intelligence.

From this perspective, reasoning is essentially the logical composition of knowledge. DeepSeek R1 and OpenAI’s o1 embody a methodology called “deep thinking,” which builds long reasoning chains (GRPO) by combining large amounts of knowledge through deduction, induction, and reflection.

Looking ahead, Wu outlines two promising research avenues: (1) the evolution of reasoning architectures toward more efficient, low‑energy models, akin to the transition from RNNs to Transformers and Mixture‑of‑Experts; and (2) the exploitation of deep‑thinking processes themselves, which can generate new knowledge by recombining existing information, though challenges remain in validating the usefulness of such compositions and achieving domain generalization. Recent efforts such as CodePlan, PromptCoT, and AlignXplore aim to address these challenges.

In summary, the advent of DeepSeek R1 opens a new door for machine reasoning, and the journey toward generalized artificial intelligence is just beginning.

Author: Wu Wei, Head of Natural‑Language Processing, Ant Financial Technology Research Institute