No Fine-Tuning Needed: How Models Actively Use Context for Continuous Learning

​About the Talk

​As large language models (LLMs) support longer context windows, long context has become a powerful but underused capability.

​Most approaches treat long context as passive input expansion. This leads to two problems:

• ​Context cannot accumulate experience across tasks

• ​Longer context increases cost without stable performance gains

​In this talk, we introduce Agentic Context Engineering (ACE), a new paradigm that treats context as an evolving learning and memory system.

​ACE uses a generate–reflect–curate loop to:

• ​Identify high-value experiences

• ​Turn them into reusable context

• ​Enable learning during inference

• ​Improve performance without updating model parameters

​We evaluate ACE across multiple agent tasks, including long-horizon execution and cross-task transfer. Results show consistent performance improvements with better efficiency.

​Paper:
Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models (ICLR 2026)
https://arxiv.org/abs/2510.04618

​Speaker: Qizheng Zhang
PhD Candidate, Stanford University (https://alex-q-z.github.io/)
Research: continual learning and self-improving AI systems

​Host: Wujiang Xu
PhD Candidate, Rutgers University
Research: LLM agents and reinforcement learning