SMITH: Small AI Models Learn to Build Their Own Tools

Appier's NeurIPS-accepted SMITH framework lets a 4B-parameter model build tools that beat a 30B model's, cutting repeated reasoning tokens about 32-fold.

@sharedot

SMITH: Small AI Models Learn to Build Their Own Tools

By @sharedot · · 7 pages

  • AI Frontier
  • AI Agents
  • Neurips
  • Tool Use

Appier's NeurIPS-accepted SMITH framework lets a 4B-parameter model build tools that beat a 30B model's, cutting repeated reasoning tokens about 32-fold.

What happened

Appier (TSE: 4180) announced that its paper "Joint Optimization of Tool Creation and Use for Large Language Model Agents" was accepted at NeurIPS.

The paper introduces SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that puts tool building and tool use into one training loop, so a model learns to build good tools and call them well at the same time.

Why it's surprising

Existing methods typically assign tool creation and use to separate models, so the tool-builder gets little feedback on real-world performance.

PR Newswire reports that in Appier's study, a model of about 4 billion parameters trained with SMITH built tools that outperformed every other method on unseen tasks — even beating a roughly 30-billion-parameter model building tools on the fly.

Effective tool creation, it turns out, doesn't require bigger models.

The evidence

According to PR Newswire, SMITH trains from 4 simple examples and then tests on 16 harder, previously unseen problems, keeping only tools that solve new tasks in a shared library.

Tools built by the small model worked even when used by a lightweight model of only about 350 million parameters, and repeated reasoning fell from 3,206 tokens to about 100 — roughly a 32-fold efficiency gain.

The stakes

The practitioner playbook is shifting toward agents that combine models, tools, memory, and guardrails into reliable systems, as the Bitcoin Foundation's 2026 agent-building guide details.

SMITH targets a persistent pain point in that workflow: tools that are poorly described or fail when other models call them. Because SMITH trains on descriptions and parameter specs rather than underlying code, clarity of documentation becomes direct training feedback, per PR Newswire.

What comes next

Appier CEO Dr. Chih-Han Yu framed the direction as agents evolving the way humans do: turning problem-solving experience into reusable tools so they never start from scratch.

The company says proven tools can be shared across agents and model sizes, making multi-agent collaboration more efficient, with enterprise applications from financial metric conversion to ad buying where repeated reasoning becomes a callable, shared tool.

Sources

  1. prnewswire.com › Appier Research Accepted at NeurIPS: AI Agents Learn Not Only to Use Tools, but to Build Their Own
  2. bitcoinfoundation.org › How to Build AI Agents in 2026: Create Your First Autonomous AI Step by Step

More on AI Frontier