Topic

Towards Experiential Intelligence: From Context Engineering to Context Learning

With the continuous enhancement of large model capabilities, agents are becoming a core form of applied intelligence. However, achieving true experiential intelligence requires more than manually curated context (Context Engineering); it demands that models learn, accumulate, and reuse experience from real interaction data—moving toward Context Learning. This talk will present Mind Lab’s engineering practices in this area. We developed **MinT**, a LoRA-RL training foundation for post-training trillion-parameter models, enabling high-throughput, low-cost reinforcement learning iterations. Using the Macaron model training as a case study, we will demonstrate how Context Learning teaches models to operate dynamic UIs, turning interaction experience into reusable model capabilities and training pipelines. Project: [https://github.com/MindLab-Research/mindlab-toolkit](https://github.com/MindLab-Research/mindlab-toolkit) Key topics include: Large Model Post-Training Evolution: From Context Engineering to Context Learning * Compare the two capability-enhancement paths and explain why, in agent scenarios, learning from interaction is critical for achieving experiential intelligence. Core Challenges of Context Learning * Discuss the difficulties in real interactions: constructing effective training signals, modeling long-term memory, and ensuring experience reusability. MinT: LoRA-RL Training Foundation for Trillion-Parameter Models * Present MinT’s core design and engineering implementation, showing how it achieves high-throughput, low-cost post-training and supports rapid iteration in agent and tool-use scenarios. Macaron Case Study: Training Dynamic UI Interaction via Context Learning * Using the Macaron model as an example, illustrate how experience trajectories are extracted from user interaction data, embedding multi-step UI interaction capabilities into model parameters and forming an iterative training pipeline.