TSCG: Deterministic Tool-Schema Compilation Boosts LLM Tool-Use Accuracy

The newly developed deterministic compiler, TSCG, addresses the discrepancies between JSON tool schemas and small language models with 4B-14B parameters. Frameworks for production agents such as OpenAI Function Calling, Anthropic Tool Use, and MCP utilize JSON for tool schemas, which, while optimized for machine parsing, are not effectively interpreted by LLMs, leading to frequent tool-use failures in production catalog sizes. TSCG transforms JSON schemas into token-efficient structured text at the API boundary, eliminating the need for model access, fine-tuning, or runtime searches. It employs eight composable operators and guarantees a formal compression bound of at least 51% for well-formed schemas. In the TSCG-Agentic-Bench benchmark, which includes around 19,000 calls across 12 models and 5 scenarios, TSCG improved Phi-4 14B’s accuracy from 0% to 84.4% with 20 tools and 90.3% with 50 tools. It also achieved a 108-181% accuracy-retained ratio across three models on BFCL. The format-versus-compression decomposition (R²=0.88 → 0.03) indicates that representation is the key factor.