Zhejiang U & Baidu's Agent-R²: 4B agent beats larger open-source baselines on GDPval
ZhejiangUniversity · hf · 2026-11-13
Zhejiang University and Baidu released Agent-R², a real-source data synthesis framework for scaling agent training on realistic productivity work (paper and model on Hugging Face).
- Problem: agents still struggle with long-horizon professional workflows, heterogeneous file contexts, and producing usable deliverables — bottlenecked by lack of scalable post-training data reflecting real work intent, authentic file environments, and deliverable-level evaluation
- Method: instead of manually specified workflows or LLM-generated files, synthesis is grounded in real occupational seeds and corpus-derived workflow evidence; a 4k-instance dataset trains Qwen3-4B-Thinking via SFT plus rubric-based RL
- Results: +18.6 points on GDPval Pass@1 to 51.0%, surpassing DeepSeek and Qwen3-235B among open-source baselines; closed-source GLM 4.7 (61.1) and Claude Opus 4.5 (72.6) still lead
More from coding & agent
- HiVe-Mem: NTU-led framework for multimodal agent long-term memory — NanyangTechnologicalUniversity · 2026-11-13
- ArtCraft clones Photoshop with AI as open source, has shipped 17 tools and wants to 'rebuild everything' — altryne · 2026-11-13
- SOC practitioners ask: AI security tools keep adding queues, not answers — RileyNet44 · 2026-11-13
- When playing and making games become the same activity: agents rewrite game dev economics — TAbrodi · 2026-11-13
- Firecrawl CTO on growing a 1,000-star repo to 1.2M developers — garrytan · 2026-11-13
- Swapping one month of Scaleway bills for a 16GB ThinkCentre to run agents locally — jiayuan_jy · 2026-11-13