Researchers have introduced ParaTool, a novel framework designed to enhance large language models' (LLMs) ability to utilize external tools. Unlike traditional methods that embed tool documentation within the model's context, ParaTool projects each tool into a distinct set of parameters. This approach aims to reduce inference overhead and minimize hallucination risks associated with lengthy contexts. The framework involves parametric tool pre-training, soft tool selection via a gating network, and joint fine-tuning of tool parameters. Experiments on benchmarks like Stable ToolBench and BFCL show ParaTool outperforming in-context learning baselines while decreasing computational complexity. AI
IMPACT This new framework could reduce computational costs and improve the reliability of LLMs when interacting with external tools.
RANK_REASON Academic paper detailing a new framework for LLM tool use. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
