Internalizing Tool Knowledge in Small Language Models via QLoRA Fine-Tuning

In three linesResearchers demonstrate that small models (Gemma 4 E4B, Qwen3-4B) fine-tuned with 8-bit QLoRA internalize tool knowledge without requiring tool schemas in prompts. On AssetOpsBench, fine-tuned models outperform unfine-tuned baselines: 82.6% input length reduction, AT-F1 of 0.65 vs 0.47, and 2.5× faster for Qwen3.

Summary generated by Claude — human-verified