简体中文 · English

SMITH: Joint training for creating and using tools¶

Core mechanism mini-suite. The same agent alternates build and use, with independent schema, code and outcome verifiers.

Paper information¶

Field	Details
Paper	arXiv 2608.24571
Institution	Appier AI Research / National Taiwan University (first author affiliations)
First public release	2026-08-25 (arXiv v1)
Author code	Yes: appier-research/smith.
Local adapter or method key	`smith`
Local reproduction code	`src/auto_research/agent_research/latest_20260826.py`

Original paper summary¶

Background and main changes¶

Tool creation is often done at inference with a frozen model, without joint feedback for the builder and user. SMITH mixes build tasks that write schemas/backends from examples with held-out tasks that use the tool pool. It rewards schema validity, code execution and final outcomes separately, with harder questions encouraging reusable abstractions.

flowchart LR
  E[few-shot examples] --> B[build schema + backend]
  B --> P[tool pool]
  P --> U[held-out use task]
  B --> S[schema reward]
  B --> C[code reward]
  U --> O[outcome reward]
  S --> R[joint RL policy]
  C --> R
  O --> R

Key figure from the paper¶

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ \mathcal J(\theta)=\mathbb E_{\tau\sim\pi_\theta}[r_{schema}(\tau)+r_{code}(\tau)+r_{outcome}(\tau)], \]

\[ \tau\sim\operatorname{Mix}(\mathcal D_{build},\mathcal D_{use}). \]

Offline and online results¶

Qwen3-4B reports 79.8% macro accuracy on held-out Reasoning-Gym, 40.4 on TabMWP-Hard and 42.6 on GQA. GQA is +7.6 points over an inference-time baseline with the same backbone. No industrial A/B.

Local reproduction¶

PlanBench-mini, 120 episodes: 12 tools built, 108 reuses and 360 verifier calls across the three dimensions. Joint success 1.0000, average cost 0.4550. See metrics and batch index.

auto-research agent-eval --method smith --benchmark planbench-mini --episodes 120 --seed 42

Scope¶

Local tools are structured plans. Build/use state and all three verifier dimensions execute separately. Qwen3-4B, an executable Python sandbox, Reasoning-Gym, TabMWP and GQA were not run.