New SMITH framework lets small models rival larger ones with far fewer tokens, paving the way for scalable multi-agent collaboration

SINGAPORE, Sept. 30, 2026 /PRNewswire/ -- Appier (TSE: 4180), an AI-native company delivering Agentic AI as a Service (AaaS), today announced its latest research paper, "Joint Optimization of Tool Creation and Use for Large Language Model Agents," accepted at NeurIPS, the world's premier AI and machine learning conference, often called "the Olympics of AI." The paper introduces SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that lets AI build tools and use them effectively in a single training loop, with each tool continuously refined based on real-world problem-solving results. The breakthrough tackles a key challenge in Agentic AI: models can create tools, but often struggle to use them well.

The research shows that small models trained with SMITH can build reusable tools that rival those created by much larger models, even on tasks they have never seen. These tools can also be shared across models and tasks, significantly reducing the token cost of repeated reasoning. The paper's acceptance at NeurIPS highlights Appier's research strength in improving AI agent efficiency and advancing multi-agent collaboration, and reinforces its position at the forefront of global Agentic AI research.

"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," said Dr. Chih-Han Yu, CEO and Co-founder of Appier. "This research shows that agents can learn to build tools, continuously refine them, and share proven tools across models of all sizes, making multi-agent collaboration more efficient and scalable. NeurIPS's acceptance of this paper further recognizes Appier's forward-looking research and innovation. We will continue to bring Agentic AI into real-world applications and deliver measurable results for businesses."

AI agents must build tools and know whether they work

As Agentic AI handles more enterprise workflows on its own, selecting and calling the right external tools has become critical for real-world deployment. Yet many AI systems still rely on engineers to build APIs or configure fixed tools in advance, which must be rebuilt whenever data sources, tasks, or business needs change. Even when AI can create its own tools, existing methods typically assign creation and use to separate models. As a result, the tool-building model gets little feedback on real-world performance, and cannot easily tell whether a tool is clearly described, works reliably, or can be called correctly by other models.

SMITH puts both skills into one training loop, so the AI learns to build good tools and use them well at the same time. When a tool has a vague description, poorly designed parameters, or fails to run, that result feeds back into the model. A closed loop of building, using, verifying, and refining keeps improving tool quality. The study found that training tool creation and tool use together clearly beats training them separately.

"When SMITH trains a model to use tools, it sees only the tool's description and parameter specifications, not the underlying code," said Chieh-Yen Lin, Research Scientist at Appier. "This makes the clarity of each description, and whether the tool can be called correctly, direct feedback during training. Our experiments also confirmed that other models can use these tools to solve problems more effectively. Looking ahead, we hope to build models that can continuously interact with their environment and take on a wider range of tasks."

Learning from 4 examples, proven on 16 new problems

SMITH trains from easy to hard. The AI first learns a method from 4 simple examples and builds tools from it. It then tests the model on 16 harder, previously unseen problems to see whether it can generalize. It keeps only tools that solve new problems and adds them to a shared tool library that multiple AI agents can use. As more high-quality tools come in, better ones replace weaker ones. Instead of building new tools for every task, agents can build and reuse proven problem-solving skills.

The research delivers three key findings:

1. Small models can outperform larger ones at tool creation: A model of about 4 billion parameters trained with SMITH built tools that outperformed those from every other method in the study on unseen tasks. It even beat a baseline in which a roughly 30-billion-parameter model built tools on the fly. Effective tool creation doesn't have to depend on bigger models.
   
   
   
   
2. Proven tools work across model sizes: Tools built by the small model handled new tasks well even when used by a lightweight model of only about 350 million parameters. The same tools also boosted the performance of larger models, allowing agents to divide work more flexibly and efficiently.
   
   
   
   
3. Repeated reasoning becomes a reusable tool: SMITH turns repeated reasoning into tools that can be called directly. In experiments, average output fell from 3,206 tokens with conventional step-by-step reasoning to about 100 tokens, roughly a 32-fold gain in efficiency. When facing similar problems, the AI doesn't need to run the full reasoning process every time, maintaining task performance while speeding up reasoning and lowering compute costs.

This research opens a new direction for enterprise Agentic AI. Many daily operations are repetitive, from converting financial metrics and processing data to querying reports, checking rules, and routing customer service cases. AI could turn these scattered methods into proven, shared tools that any agent can call, cutting the cost of repeated development and reasoning.

In advertising and marketing, agents handling customer data, personalization, customer service, and ad buying can share proven tools and consistent business rules to collaborate more efficiently. Whether entering a new market, onboarding a new advertiser, or starting with limited data, businesses can turn past successes into verifiable, scalable AI capabilities that help agents adapt to new tasks faster. Appier will continue to advance Agentic AI through forward-looking research, making AI a core engine of long-term business growth.

About Appier

Appier (TSE: 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 the vision of "Making AI Easy by making software intelligent," Appier endeavors to help businesses turn AI into ROI with its Ad Cloud, Personalization Cloud, and Data Cloud solutions. Now Appier has 17 offices across APAC, the US and EMEA, and is listed on the Tokyo Stock Exchange. Visit www.appier.com for more company information, and visit ir.appier.com/en/ for more IR information.

View original content to download multimedia:https://www.prnewswire.co.uk/news-releases/appier-research-accepted-at-neurips-ai-agents-learn-not-only-to-use-tools-but-to-build-their-own-302894483.html

Eine Podiumsdiskussion der Wiener FPÖ zum Titel „Der Preis des Klimakommunismus – Wer zahlt die Zeche?“ sorgt in der Stadtpolitik für erheblichen Wirbel. Die Freiheitliche Akademie Wien, das Freiheitliche Bildungsinstitut und der FPÖ-Rathausklub laden am 6. Oktober in die Volkshalle des Wiener Rathauses und holen dafür James Taylor, Präsident des US-amerikanischen Heartland Institute, nach Wien. Taylor gilt seit Jahren als prominenter Akteur der internationalen Szene, die den menschengemachten Klimawandel in Zweifel zieht, und steht seit Langem in der Kritik, weil seine Organisation nachweislich Geld aus der fossilen Industrie erhalten hat.

Auf dem Podium sollen neben Taylor der Wiener FPÖ-Chef Dominik Nepp, Gerald Zmuegg von der Freiheitlichen Wirtschaft Wien sowie der Unternehmer und Politblogger Gerald Markel diskutieren. Eröffnet wird der Abend vom Wiener FPÖ-Klubobmann Maximilian Krauss. Inhaltlich bedient die FPÖ damit ein Narrativ, das sie seit geraumer Zeit in der Klimadebatte verwendet: Klimaschutzmaßnahmen werden als „Klimakommunismus“ adressiert und damit als ideologisch motivierte, wirtschaftlich schädliche Eingriffe gerahmt. Im Zentrum der Veranstaltung steht laut Ankündigung nicht die Abrechnung mit den Kosten der aktuellen Klimaauswirkungen, sondern mit jenen einer ambitionierten Klimapolitik.

Die Einladung Taylors fällt in ein Jahr, das in Europa und Österreich von außergewöhnlichen Wetterextremen geprägt ist. 2026 brachte neue Temperaturrekorde, eine hohe Zahl an Hitzetagen und Tropennächten und eine sich seit dem Frühjahr verschärfende Trockenheit. In der Landwirtschaft werden die Hitzeschäden europaweit auf 14 bis 17 Milliarden Euro geschätzt, für Österreich beziffert die Hagelversicherung den Dürreschaden auf rund 1 Milliarde Euro. Die Agentur für Gesundheit und Ernährungssicherheit (AGES) meldet zudem eine Rekordzahl an Hitzetoten: Im Juni und Juli starben demnach 646 Menschen in Österreich an den Folgen der Hitze. Vor diesem Hintergrund kritisieren Gegner der Veranstaltung, die FPÖ ignoriere die realen Kosten und Risiken der Erderwärmung, während sie über vermeintliche Kosten von Klimapolitik diskutieren lässt.

Wiens Bürgermeister Michael Ludwig (SPÖ) distanzierte sich in einer Stellungnahme deutlich von der inhaltlichen Ausrichtung des Abends. Die Stadt bekenne sich zu den wissenschaftlichen Erkenntnissen über den menschengemachten Klimawandel und zu einer verantwortungsvollen Klimapolitik, betonte er. Positionen, die Klimaschutz pauschal als „Klimakommunismus“ diffamieren, lehne er ausdrücklich ab. Zugleich verwies Ludwig auf die Rolle des Rathauses als „Haus der Demokratie“: Den im Gemeinderat vertretenen Fraktionen stünden grundsätzlich Räumlichkeiten für Veranstaltungen zur Verfügung. Kritik kam auch von den Wiener Grünen, deren Vorsitzender Peter Kraus die Einladung Taylors als „Verhöhnung“ der Bevölkerung bezeichnete, die mit den Folgen von Hitze und Dürre konfrontiert sei, und den Bürgermeister aufforderte, zu prüfen, ob für eine derartige Veranstaltung tatsächlich Räume im Rathaus zur Verfügung gestellt werden sollten.