Appier's SMITH trains AI agents to forge their own tools — a 4B model beat a 30B one at it

Appier announced today that its paper “Joint Optimization of Tool Creation and Use for Large Language Model Agents” has been accepted at NeurIPS. The paper introduces SMITH — Schema-grounded Multi-task Iterative Tool Honing — a reinforcement-learning framework that trains an AI agent to both create tools and use them inside a single training loop.

The problem it attacks is the one every agentic deployment eventually hits: models are only as useful as the tools they can call, and someone — usually an engineer — has to build those tools. Existing tool-creation methods hand the job to a separate model from the one that uses the result, so the builder never gets feedback on whether a tool is clearly described, runs reliably, or can actually be invoked. SMITH closes that loop: when a tool’s description is vague, its parameters misdesigned, or its code fails, the failure feeds straight back into training.

Build it, use it, keep what works#

SMITH trains easy to hard. The agent first learns a method from 4 simple examples and writes tools from them, then faces 16 harder problems it has never seen. Only tools that solve new problems survive, joining a shared library that multiple agents can draw on — and better tools replace weaker ones over time. Each rollout is either a build task (write a tool from a few examples) or a use task (invoke a pooled tool on a held-out question), with three separate reward axes catching schema, code, and outcome failures independently, so each failure mode teaches its own lesson.

A 4B model out-built a 30B one#

Three results stand out. First, a Qwen3 model of about 4 billion parameters trained with SMITH wrote tools that outperformed every other method in the study on unseen tasks — including a baseline in which a roughly 30-billion-parameter model built tools on the fly. On 13 procedural-reasoning tasks with exact verifiers, the 4B model reached 79.8% macro-average accuracy, the best score in the evaluation. Second, the tools travel: written by the small model, they lifted a 350-million-parameter lightweight model and also boosted larger models — which makes it cheaper to divide work across agents of different sizes.

The 32× efficiency play#

Third is the efficiency number. Because SMITH compiles repeated reasoning into callable tools, average output fell from 3,206 tokens with conventional step-by-step reasoning to about 100 — roughly a 32-fold drop — while task performance held. For anyone running agents at scale, that is the difference between a rounding error and a budget line.

“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, Appier’s CEO and co-founder. “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.”

Chieh-Yen Lin, a research scientist at Appier, described the mechanism: “When SMITH trains a model to use tools, it sees only the tool’s description and parameter specifications, not the underlying code.” That constraint turns description clarity into direct training feedback.

What to watch#

Some proportion is in order. This is a company-announced research result — Appier is a Tokyo-listed adtech and martech vendor (TSE: 4180) — and the numbers come from 13 procedural-reasoning tasks with exact verifiers, a controlled setting rather than the open web. A NeurIPS acceptance is peer review’s stamp of interest, not a product launch. Still, the direction matches where the whole agentic stack is heading: tools agents write for themselves, verified by execution, shared across fleets. Appier frames the payoff in enterprise terms — repetitive operations like converting financial metrics, processing data, querying reports, and routing support cases could stop being rebuilt by hand every time and become proven, callable tools instead.