CATS: A Model-First Notation for Compact LLM Tool Definitions
Year: 2026
Type: preprint
Abstract: Modern large language models (LLMs) are given tool definitions in JSON Schema placed in the model's input. Yet this approach has a key issue: JSON Schema is verbose, because it was designed for machine parsers and not LLMs. Hence, passing JSON Schema definitions to LLMs spends tokens on grammar and structure that the LLM does not need — thereby increasing costs for developers deploying these models. We introduce CATS (Compact Agent Tool Schema), a notation designed to communicate tool definitions in a more token-efficient manner. When converting from JSON Schema to CATS, CATS converts losslessly by design: rather than encoding constructs imprecisely, it falls back to raw JSON Schema for anything it cannot represent without changing behavior, making adoption safe. Across 3105 deduped real tool definitions from the Berkeley Function Calling Leaderboard (BFCL), CATS only had to fall back a total of 5 times for certain edge cases (99.84% coverage). Across the 3100 convertible schemas, CATS reduces per-tool input tokens by a pooled median of 29.9% across three tokenizers, and never exceeds the token count of the original JSON Schema. In a pre-specified non-inferiority test over three models with reasoning disabled, tool-calling accuracy is largely preserved, although effects vary by model. GPT-5.4 satisfies the pre-specified 3-percentage-point non-inferiority criterion, while Claude Sonnet 4.6 and Qwen3.5-35B-A3B yield inconclusive results because their confidence intervals cross the pre-specified threshold. We release the specification, a reference converter, a deployment protocol, and the full evaluation harness.
Author: Mahajan, Armaan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225461049