OpenThoughts-Agent: Data Recipes for Agentic Models

Negin Raoof, Kelly Buchanan, Artem Gazizov, Reinhard Heckel, Chinmay Hegde, Sankalp Jajee, Daanish Khazi, Emmanouil Koukoumidis, Xiangyi Li, Hange Liu, Shlok Natarajan, Harsh Raj, Nicholas Roberts, Ethan Shen, Nishad Singhi, Michael Siu, Ashima Suvarna, Hanwen Xing, Patrick Yubeaton, Robert Zhang, Leon Liangyu Chen, Xiaokun Chen, Steven Dillmann, Saadia Gabriel, Xunyi Jiang, Anurag Kashyap, Boxuan Li, Yein Park, Minh Pham, Sujay Sanghavi, Lin Shi, Ke Sun, Yixin Wang, Zhiwei Xu, Erica Zhang, Siyan Zhao, Wanjia Zhao, Jenia Jitsev, Alex Dimakis, Benjamin Feuer, Ludwig Schmidt (approx. 50 authors total)

Open source

Type: paper
Venue: arXiv (cs.AI)
Year: 2026
Source: x
Access: public
Language: en
Added: 2026-09-29
Verified: 2026-09-29

Summary

Addresses the lack of public knowledge on data curation for agentic models via 100+ controlled ablations on a six-stage SFT curation pipeline plus agentic RL data studies. Fine-tuning Qwen3-32B on the resulting 100K-example OpenThoughts-Agent-v2 yields 44.8% average across seven agentic benchmarks (+3.9pp over the strongest open-data baseline), with strong data scaling properties.

Research notes

Key findings: Task source and diversity matter most in the SFT curation pipeline; OpenThinker-Agent-32B is the strongest open-data <=32B model on the seven-benchmark average (44.8%, +3.9pp over best baseline); SFT and RL stages can compose (8B-scale evidence)
Limitations: RL study only at 8B scale; no base-model ablation (all Qwen3); largest set 100K trajectories - extrapolation to multi-million-trajectory regimes untested.
Announced by Alex Dimakis on X 2026-06-24; blog published 2026-06-10 at open-thoughts.ai/blog/agent.