ToolLoop: Three-Stage Reverse Synthesis of Tool-Call Training Data (EMNLP 2026)

jiqizhixin · x · 2026-09-29

Tool calls let LLMs reach external APIs, but synthesizing tool-use training data usually follows a "generate first, filter later" pipeline: the model emits user questions and tool calls in one shot, then rules or a model decide what to keep. Malformed calls often slip past single-pass format checks, and the pipeline optimizes for final pass/fail rather than the actual target function.

ToolLoop, accepted to EMNLP 2026 Main Conference, splits synthesis into three stages — target-function sampling, reverse user-question generation, and forward tool-call generation — with dynamic self-feedback at each stage. It uses explicit intermediate results to constrain later generation, aligning the target function, user intent, and tool calls via generate-verify-correct.

Original post →

More from Models

Models channel →