On March 26, Lin Junyang, the former technical lead of Alibaba’s Tongyi QianWen (Qwen) large language model project, published his first in-depth essay since leaving the company. Titled "From Reasoning-Based Thinking to Agent-Based Thinking", the piece appeared on a social platform, offering a systematic analysis of the reasoning model era while charting potential trajectories for AI’s evolution.

Lin highlighted that during the first half of 2025, the AI industry prioritized reasoning-based thinking, with milestones such as OpenAI’s o1 and DeepSeek-R1 demonstrating enhanced model reasoning capabilities through reinforcement learning techniques. However, he argued that this phase has now reached its objectives.

Looking ahead, Lin identified agent-based thinking as the next frontier. In this paradigm, AI models are designed to think with the explicit purpose of driving action, dynamically refining strategies based on real-time environmental feedback. This marks a departure from static reasoning toward adaptive, goal-oriented intelligence.

Lin disclosed that during his tenure at QianWen, the team explored integrating reasoning and instruction modes to enable adjustable reasoning intensity. However, this initiative ultimately failed due to fundamental mismatches in data distribution patterns and behavioral objectives between the two approaches.

He emphasized that the industry is undergoing a paradigm shift: from "training models" to "training agents". The critical question now centers not on whether models can sustain prolonged reasoning, but whether they can reason effectively to drive purposeful actions.

Lin predicted that agent-based thinking will soon dominate AI development, displacing traditional static reasoning frameworks. He stressed that even when tackling highly complex challenges, advanced systems must demonstrate capabilities across six key dimensions: searching for relevant information, simulating potential outcomes, executing plans, inspecting results, verifying accuracy, and correcting errors through iterative refinement.