AgentFrontier: Expanding the Capability Frontier of LLM Agents with ZPD-Guided Data Synthesis
Xuanzhong Chen, Zile Qiao, Guoxin Chen, Liangcai Su, Zhen Zhang, Xinyu Wang, Pengjun Xie, Fei Huang, Jingren Zhou, Yong Jiang
2025-10-29
Summary
This paper focuses on improving large language models (LLMs) by training them on tasks that are challenging but achievable with a little help, ultimately making them better at complex reasoning.
What's the problem?
LLMs often struggle with truly difficult problems that require advanced reasoning skills. Simply giving them more data doesn't always help, and it's hard to figure out *what* kind of data would actually push their abilities forward. The core issue is finding tasks that are at the edge of what the LLM can currently do – not too easy, but not impossible.
What's the solution?
The researchers developed a system called the AgentFrontier Engine. This engine automatically creates new training data specifically designed to be within an LLM’s ‘Zone of Proximal Development’ (ZPD). Think of the ZPD like a student learning something new – it’s what they can’t do *alone*, but can accomplish with guidance. The engine generates complex, multidisciplinary problems and then uses this data to both pre-train and fine-tune the LLM. They also created a benchmark, the ZPD Exam, to measure how well the LLM is improving on these challenging tasks. They then used this system to train a new model, AgentFrontier-30B-A3B.
Why it matters?
This work is important because it provides a scalable way to build more intelligent LLMs. Instead of just throwing more data at the problem, this approach focuses on *targeted* training that pushes the LLM to learn and improve its reasoning abilities. The AgentFrontier-30B-A3B model actually outperformed other leading models on difficult tests, showing that this ZPD-guided training is a promising path towards creating more capable AI agents.
Abstract
Training large language model agents on tasks at the frontier of their capabilities is key to unlocking advanced reasoning. We introduce a data synthesis approach inspired by the educational theory of the Zone of Proximal Development (ZPD), which defines this frontier as tasks an LLM cannot solve alone but can master with guidance. To operationalize this, we present the AgentFrontier Engine, an automated pipeline that synthesizes high-quality, multidisciplinary data situated precisely within the LLM's ZPD. This engine supports both continued pre-training with knowledge-intensive data and targeted post-training on complex reasoning tasks. From the same framework, we derive the ZPD Exam, a dynamic and automated benchmark designed to evaluate agent capabilities on these frontier tasks. We train AgentFrontier-30B-A3B model on our synthesized data, which achieves state-of-the-art results on demanding benchmarks like Humanity's Last Exam, even surpassing some leading proprietary agents. Our work demonstrates that a ZPD-guided approach to data synthesis offers a scalable and effective path toward building more capable LLM agents.