RAGEN-2: Reasoning Collapse in Agentic RL
Zihan Wang †∗,1 , Chi Gui †,12 , Xing Jin †,3 , Qineng Wang †,1 , Licheng Liu †,4 ,Kangrui Wang 1 , Shiqi Chen 5 , Linjie Li 6 , Zhengyuan Yang 7 , Pingyue Zhang 1 , Yiping Lu 1 , Jiajun Wu 8 , Li Fei-Fei 8 , Lijuan Wang 7 , Yejin Choi 8 , Manling Li 1 † Core contributors. ∗ Project lead. 1 Northwestern University 2 UIUC 3 Independent 4 Imperial College London 5 Oxford University 6 University of Washington 7 Microsoft 8 Stanford University https://ragen-ai.github.io/v2/ · 2026
Read RAGEN-2: Reasoning Collapse in Agentic RL by Zihan Wang †∗,1 , Chi Gui †,12 , Xing Jin †,3 , Qineng Wang †,1 , Licheng Liu †,4 ,Kangrui Wang 1 , Shiqi Chen 5 , Linjie Li 6 , Zhengyuan Yang 7 , Pingyue Zhang 1 , Yiping Lu 1 , Jiajun Wu 8 , Li Fei-Fei 8 , Lijuan Wang 7 , Yejin Choi 8 , Manling Li 1 † Core contributors. ∗ Project lead. 1 Northwestern University 2 UIUC 3 Independent 4 Imperial College London 5 Oxford University 6 University of Washington 7 Microsoft 8 Stanford University https://ragen-ai.github.io/v2/ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-06268.txt.
Description
Wang and fifteen colleagues introduce RAGEN-2, a study of reasoning collapse in reinforcement-learning training of multi-turn agents. Even with stable entropy, which measures only within-input diversity, models can lean on set templates that seem diverse but ignore inputs. This is template collapse, invisible to entropy and existing metrics. Reasoning quality splits into within-input diversity and cross-input distinguishability. Across tasks, mutual information tracks final performance far more strongly than entropy. SNR-Aware Filtering selects high-signal prompts each iteration using reward variance, improving task performance.
Catalog record
| CREATOR | Zihan Wang †∗,1 , Chi Gui †,12 , Xing Jin †,3 , Qineng Wang †,1 , Licheng Liu †,4 ,Kangrui Wang 1 , Shiqi Chen 5 , Linjie Li 6 , Zhengyuan Yang 7 , Pingyue Zhang 1 , Yiping Lu 1 , Jiajun Wu 8 , Li Fei-Fei 8 , Lijuan Wang 7 , Yejin Choi 8 , Manling Li 1 † Core contributors. ∗ Project lead. 1 Northwestern University 2 UIUC 3 Independent 4 Imperial College London 5 Oxford University 6 University of Washington 7 Microsoft 8 Stanford University https://ragen-ai.github.io/v2/ |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 103362 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.06268