Data as of Jul 25, 2026 · Based on 2,718,867 AI responses across 9,511 prompts · See how Parse measures this
SPORT-Agentssport-agents.github.ioSPORT stands for Iterative Tool Usage Exploration for Multimodal Agents via Step-wise Preference Tuning and is an online self-exploration framework that lets multimodal agents autonomously generate tasks, sample and verify candidate steps with an LLM, and update their controller through preference tuning without human annotations. It uses real executable tools across web search, vision, image editing, file understanding, multi-modal reasoning, and Python packages to solve the generated tasks. Evaluations on GTA and GAIA show SPORT agents outperforming SFT baselines and open-source models, demonstrating improved tool use, coding ability, and generalization.
Parse Score
