The new SMITH framework enables smaller models to compete with larger models using fewer tokens, paving the way for scalable multi-agent collaboration.
Singapore, September 30 /PRNewswire/ -- Appier (Tokyo Stock Exchange: 4180) today announced that its latest research paper, "Joint Optimization of Tool Creation and Use for Large Language Model Agents," has been accepted by NeurIPS. NeurIPS is a premier global conference on artificial intelligence and machine learning, often referred to as the "Olympics of AI."
This paper introduces the SMITH (Schema-grounded Multi-task Iterative Tool Honing) framework, a reinforcement learning framework that enables AI to build and effectively use tools within the same training loop, with each tool continuously optimized based on real-world problem-solving outcomes. This breakthrough addresses a key challenge in agentic AI: models can create tools but often struggle to use them effectively.
Research shows that small models trained with SMITH can build reusable tools whose performance rivals that of tools created by larger models—even on tasks the models have never seen. These tools can also be shared across different models and tasks, significantly reducing the token cost of redundant reasoning. The paper has been accepted by NeurIPS, highlighting Appier’s research strength in enhancing AI agent efficiency and advancing multi-agent collaboration, while reinforcing its position at the forefront of global Agentic AI research.
Dr. Chou Hsi-han, Co-founder and CEO of Appier, said: “Humans transform their problem-solving experiences into tools, so they don’t have to start from scratch each time. AI agents are now evolving in the same way. This study demonstrates that agents can learn to build tools, continuously optimize them, and share proven tools across models of different scales, making multi-agent collaboration more efficient and scalable. The acceptance of this paper by NeurIPS further validates Appier’s forward-looking research and innovation. We will continue to bring Agentic AI into real-world applications and deliver measurable outcomes for enterprises.”
AI agents must be able to build tools and determine whether those tools are effective.
As Agentic AI increasingly autonomously handles enterprise workflows, selecting and invoking the right external tools has become critical for real-world deployment. However, many AI systems still rely on engineers to pre-build APIs or configure fixed tools, which must be rebuilt whenever data sources, tasks, or business requirements change. Even when AI can create tools itself, existing approaches typically assign "creation" and "usage" to separate models. As a result, the model responsible for building tools receives almost no feedback on real-world performance and struggles to determine whether the tool descriptions are clear, the execution is reliable, or other models can correctly invoke them.
SMITH combines these two capabilities into a single training loop, enabling the AI to simultaneously learn how to build effective tools and how to use them well. When tool descriptions are vague, parameter designs are suboptimal, or operations fail, these outcomes are fed back into the model. The cycle of building, using, validating, and optimizing forms a closed loop that continuously improves tool quality. Research shows that training tool creation and tool usage together significantly outperforms training them separately.
Appier research scientist Lin Zheyan said: “When the SMITH training model uses tools, it only sees the descriptions and parameter specifications of the tools, not the underlying code. This ensures that whether each description is clear and whether the tools can be correctly invoked generates direct feedback during training. Our experiments have also confirmed that other models can use these tools more effectively to solve problems. Looking ahead, we aim to build models that can continuously interact with their environment and take on a broader range of tasks.”
Learn from 4 examples, validate on 16 new questions
SMITH employs a progressive training approach. The AI first learns a method from four simple examples and builds a tool based on it, then tests the model’s ability to generalize on 16 more difficult, previously unseen problems. The system retains only those tools capable of solving the new problems and adds them to a shared tool library for use by multiple AI agents. As high-quality tools are continuously added, stronger tools replace weaker ones. Agents no longer need to rebuild tools for each task—they can instead construct and reuse validated problem-solving capabilities.
This study presents three key findings:
- Smaller models can outperform larger models in tool creation: a model with approximately 4 billion parameters, trained using SMITH, outperforms all other methods studied in constructing tools on unseen tasks. It even surpasses the baseline of a model with approximately 30 billion parameters constructing tools on the fly. Effective tool creation does not necessarily rely on larger models.
- Verified tools work across model scales: tools built on small models perform just as well on lightweight models with only about 350 million parameters when handling new tasks. These tools also enhance the performance of larger models, enabling agents to divide tasks more flexibly and efficiently.
- Repeated reasoning can be transformed into reusable tools: SMITH converts repeated reasoning into directly callable tools. In experiments, the average output decreased from 3,206 tokens with traditional step-by-step reasoning to approximately 100 tokens, achieving a 32-fold efficiency gain. When encountering similar problems, the AI no longer needs to fully execute the reasoning process each time, accelerating inference and reducing computational costs while maintaining task performance.
This research opens new possibilities for enterprise agentic AI. Many daily operations are repetitive, such as converting financial metrics, processing data, querying reports, checking rules, and routing customer service tickets. AI can transform these disparate methods into validated, shareable tools that any agent can invoke, reducing the cost of redundant development and repeated reasoning.
In the advertising and marketing industry, agencies handling customer data, personalization, customer service, and ad buying can share validated tools and consistent business rules to collaborate more efficiently. Whether entering new markets, onboarding new advertisers, or starting with limited data, businesses can transform past successes into verifiable, scalable AI capabilities that help agencies adapt faster to new tasks. Appier will continue advancing Agentic AI through forward-looking research, making AI a core engine for long-term business growth.
About Appier
Appier (Tokyo Stock Exchange: 4180) is an AI-native Agentic AI as a Service (AaaS) company that empowers business decision-making with cutting-edge AdTech and MarTech solutions. Founded in 2012 with a vision of “Making Software Smart, Making AI Simple,” Appier helps enterprises turn AI into ROI through its Ad Cloud, Personalization Cloud, and Data Cloud solutions. Appier currently operates 17 offices across Asia-Pacific, the United States, and EMEA, and is listed on the Tokyo Stock Exchange. For more company information, visit www.appier.com; for investor relations, visit ir.appier.com/en/.