Boosting LLM Function Call Capabilities: From Data Construction to RLHF Optimization

On July 12, 2025, the DataFun Summit will host a session on enhancing large language model (LLM) Function Call capabilities, presented by Yao Yitong, a semantic algorithm engineer at China Telecom AI Research Institute.

Talk title: LLM Fundamentals Function Call Capability Enhancement: From Data Construction to RLHF Optimization Loop

Abstract: Function Call (FC) is crucial for LLM deployment and progression to agent‑level applications. Current models struggle with generalization and multi‑turn tool usage. This talk outlines systematic optimization from both data and training perspectives, covering data pattern analysis, multi‑dimensional classification, efficient data pipelines, SFT‑RLHF co‑training, evaluation metrics, and practical scenario solutions.

Outline:

Why Function Call is key for LLM adoption

Data‑side solutions: LLM FC data patterns, classification taxonomy, high‑quality data production chain

Training‑side solutions: SFT and reinforcement learning synergy, FC evaluation framework

Scenario‑specific optimization strategies

Audience takeaways: systematic understanding of FC data construction, training methods combining RLHF, and challenges in real‑world deployment

Attendees can scan the QR code to reserve a free live broadcast.