Apagear’s Research Paper Accepted to NeurIPS
TECHWORLD ·
✦ AI Summary
Apagear said its research paper, "Joint Optimization of Tool Creation and Utilization for Large Language Model Agents," has been accepted to NeurIPS.
The paper proposes the reinforcement learning framework SMITH (Schema-grounded Multi-task Iterative Tool Honing), which performs tool creation and effective utilization in a single learning loop.
The research found that small models trained with SMITH created reusable tools, and their performance was comparable to tools created by much larger models even in new tasks.
Apagear announced that its research paper, "Joint Optimization of Tool Creation and Utilization for Large Language Model Agents," has been accepted to NeurIPS. NeurIPS is an AI and machine learning conference often called the "AI Olympics."
Apagear said the paper proposes a reinforcement learning framework, SMITH (Schema-grounded Multi-task Iterative Tool Honing). The research aims to address a core challenge in agentic AI: although AI models can create tools, they struggle to use them effectively.
SMITH performs AI tool creation and effective utilization in a single learning loop. It also has a structure that continuously improves tools based on real problem-solving outcomes.
According to the research, small models trained with SMITH were able to create reusable tools. The performance of those tools in new tasks was comparable to those created by much larger models.
Tools created in advance are not limited to a specific model or task and can be reused, pointing to a way to significantly reduce the token costs of iterative reasoning.
Apagear co-founder and CEO Chihan Yu said that just as humans do not need to start over from scratch every time because they turn problem-solving experience into tools, AI agents are evolving in the same way. He added that the study demonstrated the potential for AI agents to create and improve their own tools, and proved that effective tools can be shared across models of different sizes. He also said this could improve the efficiency and scalability of multi-agent collaboration.
Chihan Yu said the paper's acceptance to NeurIPS is a renewed recognition of Apagear's leading research and innovation. He added that the company aims to expand the practical application of agentic AI and support enterprises in creating measurable results.
Source: TECHWORLD · Lee Gwang-jae
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407638
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.