Instruction Following Training For Api Tool Use Via In Context Learning

via “single-tool and multi-tool inference with api execution”

Framework for training LLM agents on 16K+ real APIs.

Unique: Integrates model inference with live API execution in a single pipeline, handling parameter construction, API calls, response parsing, and error recovery within the inference loop rather than as separate post-processing steps.

vs others: End-to-end inference pipeline eliminates manual API integration work, whereas generic LLM APIs (OpenAI, Anthropic) require separate function-calling and orchestration layers.

via “function calling and tool use via prompt engineering”

Mixtral 8x7B Instruct is a pretrained generative Sparse Mixture of Experts, by Mistral AI, for chat and instruction use. Incorporates 8 experts (feed-forward networks) for a total of 47 billion...

Unique: Instruction-tuning enables reliable tool-use through learned patterns without native function-calling APIs, allowing flexible tool specification and custom output formats via prompt engineering

vs others: Achieves 75-85% tool-use accuracy at 3x lower cost than GPT-4 function calling while maintaining flexibility to define custom tools and output formats through prompting

via “instruction-following training for api tool use via in-context learning”

* ⭐ 08/2023: [MetaGPT: Meta Programming for Multi-Agent Collaborative Framework (MetaGPT)](https://arxiv.org/abs/2308.00352)

Unique: Uses curriculum-based synthetic data generation to progressively teach LLMs API tool use, starting with simple single-API calls and progressing to complex multi-step workflows. Leverages the unified API schema to generate diverse, generalizable training examples without manual annotation.

vs others: Outperforms zero-shot prompting and generic instruction-following fine-tuning by using API-specific curriculum learning that mirrors real-world task complexity progression.