Congratulations! Nice and exciting work for data science agents from @FanNie1208 !
🚀 Excited to share that #DSGym has been accepted to ICML 2026!View quoted post on X
DSGym is a holistic, unified framework for evaluating and training data science agents with standardized abstractions and a modular architecture for adding tasks, agent scaffolds, and tools.
In this work, we:
🔍 Show that existing data science benchmarks are vulnerable to shortcuts: agents can often solve tasks without using the actual data.
📊 Release DSGym-Tasks, a curated task ecosystem that standardizes and audits representative benchmarks, filters shortcut-solvable tasks, and expands coverage with new scientific tasks.
⚡ Use DSGym for execution-grounded trajectory synthesis: with only 2K samples, we train a 4B model that outperforms GPT-4o on standardized data analysis benchmarks.
📄 Paper: arxiv.org/abs/2601.16344
💻 Code: github.com/fannie1208/DSG…
🤗 Dataset: huggingface.co/DSGym
🧵👇