#Japan #Tokyo #AI Agents #Appier #NeurIPS

Introduction

In a groundbreaking milestone for artificial intelligence, Appier recently presented its research at NeurIPS 2026, a pivotal gathering in the AI and machine learning community. The research paper, titled "Joint Optimization of Tool Creation and Use for Large Language Model Agents," sheds light on a new framework named SMITH, which stands for Schema-grounded Multi-task Iterative Tool Honing. This innovative approach allows small AI models to not only use tools effectively but also to construct and refine them in a continuous learning cycle.

The Innovation Behind SMITH

One of the main challenges in agentic AI has been the disparity between tool creation and their effective utilization. Traditionally, AI systems would either be dependent on engineers for building tools or rely on different models for their creation and usage, often leading to inefficiencies. SMITH addresses this gap by integrating both functionalities into a single training loop. Through reinforcement learning, AI agents can learn to develop tools that adapt based on their performance in real-world scenarios.

According to Appier's CEO and Co-founder, Dr. Chih-Han Yu, the evolution of AI agents mirrors the human process of tool-making: “Humans turn their problem-solving experience into tools, so they never have to start from scratch. AI agents are now evolving in the same way.” This perspective underlines the essence of SMITH’s innovation, which not only enhances the capabilities of AI agents but also streamlines the evolution of their tool-making proficiency.

Bridging the Gap Between Tool Creation and Usage

One of the standout features of the SMITH framework is its ability to provide direct feedback. When a tool is poorly designed or fails to operate correctly, that information is immediately fed back into the AI model, allowing it to learn from mistakes and iteratively improve tool quality. This closed-loop system of building, utilizing, and refining keeps enhancing tool effectiveness, leading to a significant leap in agent efficiency.

The research further revealed that even small models, equipped with SMITH, could outperform larger ones in tool creation. Specifically, a model with around 4 billion parameters was able to generate tools that exceeded the capabilities of those created by much larger models, showcasing that size is not always indicative of effectiveness in AI.

Application Impact and Real-World Benefits

So, what does this mean for enterprises? The applications of SMITH are vast and hold the potential to revolutionize various business workflows. Many daily operations, particularly in data analysis and customer service, often involve repetitive and time-consuming tasks. With SMITH, AI can convert these scattered methods into standardized tools, significantly reducing the time and resources required for tool development.

For example, AI agents managing customer data, advertising strategies, or service inquiries can now easily share proven tools across models and tasks, capitalizing on past successes. This shift towards a tool-sharing culture could lead to enhanced collaboration among different AI agents, paving the way for more efficient work methodologies.

Conclusions and Future Directions

Appier's research represents a significant step forward in the realm of Agentic AI, emphasizing the need for AI systems to be flexible and self-sufficient in their problem-solving capabilities. Going forward, the company aims to further explore this innovative path, continually refining the SMITH framework and its methodologies to foster the growth of AI as an integral component of business strategy.

As highlighted by Dr. Yu, this breakthrough at NeurIPS establishes Appier's commitment to leading the charge in AI research, driving not just technological advancements but also tangible business benefits. With Agentic AI transforming traditional operations, the landscape of how businesses harness technology is set for a profound shift towards a more autonomous and efficient future.