Targeted Preference Repair for Tool-Calling Reliability

This preprint reports a targeted DPO router-repair experiment on the Tool-Calling Reliability Benchmark. Starting from a Qwen2.5-3B-Instruct planner with LoRA adaptation, a 157-row preference dataset fixes observed first-tool routing confusions in customer support. Across seeds 11, 13, and 17, target task success improves from 0.8333 to 0.9074 and invalid tool-call rate drops from 0.1367 to 0.0719. A transfer matrix shows the gain is not uniform: ecommerce operations regresses from 1.0000 to 0.8889 first-tool accuracy, while fintech risk and developer tools remain unchanged.