RAGEN-2: Your Agent Trained with RL Might Have Stopped Thinking
Here is a scary finding from the RAGEN-2 team: after RL training, your agent looks like it is reasoning, but it is actually just filling in templates.
The paper calls it reasoning collapse. Early in training, the agent produces diverse, input-specific chains of thought. But as RL continues, the reasoning degrades into fixed patterns that look different on the surface but carry zero real signal across different inputs. The authors decompose this into two measurable axes: within-input diversity (are the reasoning traces varied?) and cross-input distinguishability (can you tell which input produced which reasoning?). When the second one drops to near zero, the agent is in what they call the Echo Trap, repeating the same template regardless of what it sees.
The fix is surprisingly elegant. SNR-Aware Filtering selects training prompts based on reward variance, a lightweight proxy for which examples actually require real reasoning versus which ones can be solved by pattern matching. Applied across planning tasks, math reasoning, web navigation, and code execution, the method restores genuine reasoning without sacrificing task performance.
This matters because every company fine-tuning agents with RL right now could be creating agents that pass benchmarks by memorizing templates rather than actually thinking. The RAGEN framework at 2.6K GitHub stars includes 10 built-in environments for testing, and the team behind it includes Li Fei-Fei and Yejin Choi, two of the most influential names in AI research.
https://github.com/mll-lab-nu/RAGEN
← Back to all articles
The paper calls it reasoning collapse. Early in training, the agent produces diverse, input-specific chains of thought. But as RL continues, the reasoning degrades into fixed patterns that look different on the surface but carry zero real signal across different inputs. The authors decompose this into two measurable axes: within-input diversity (are the reasoning traces varied?) and cross-input distinguishability (can you tell which input produced which reasoning?). When the second one drops to near zero, the agent is in what they call the Echo Trap, repeating the same template regardless of what it sees.
The fix is surprisingly elegant. SNR-Aware Filtering selects training prompts based on reward variance, a lightweight proxy for which examples actually require real reasoning versus which ones can be solved by pattern matching. Applied across planning tasks, math reasoning, web navigation, and code execution, the method restores genuine reasoning without sacrificing task performance.
This matters because every company fine-tuning agents with RL right now could be creating agents that pass benchmarks by memorizing templates rather than actually thinking. The RAGEN framework at 2.6K GitHub stars includes 10 built-in environments for testing, and the team behind it includes Li Fei-Fei and Yejin Choi, two of the most influential names in AI research.
https://github.com/mll-lab-nu/RAGEN
Comments