TRACE: Business Rule-Grounded Reasoning Curriculum for Knowledge-Preserving Parametric Tool Retrieval in Enterprise LLMs

TRACE is a two-stage curriculum for enterprise LLMs that preserves tool knowledge while improving retrieval accuracy.

The first stage seeds tool knowledge with LoRA, then the model is trained to emit a reasoning trace plus tool-token JSON.

It maintains earlier QA and MCQ performance while enabling low-latency single-beam greedy decoding.

Across an 8,300+ tool enterprise catalog, TRACE also improves tool understanding and retrieval recall.