TRACE: Business Rule-Grounded Reasoning Curriculum for Knowledge-Preserving Parametric Tool Retrieval in Enterprise LLMs
TRACE trains tool-token retrieval to keep API knowledge intact while cutting decoding down to greedy search.
The paper says earlier parametric retrieval methods lost tool knowledge during training and were too slow for real-time use. TRACE uses a two-stage curriculum: first seeding tool knowledge with LoRA, then training reasoning traces before JSON tool-token output. On an 8,300-plus-tool enterprise catalog, it improved MCQ and QA probing over Stage 1 and reported recall of about 86% on one domain and 60% on another. HF Daily Papers' note
The paper says earlier parametric retrieval methods lost tool knowledge during training and were too slow for real-time use. TRACE uses a two-stage curriculum: first seeding tool knowledge with LoRA, then training reasoning traces before JSON tool-token output. On an 8,300-plus-tool enterprise catalog, it improved MCQ and QA probing over Stage 1 and reported recall of about 86% on one domain and 60% on another. HF Daily Papers' note
score 5