ReTerm Enables Recursive Self-Improvement for Code Agents via Failure Recovery

1 reports archived1 independent sourcesupdated 11/4/2026, 12:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers from Fudan University and others introduce ReTerm, a framework enabling code agents to recover from their own failures and learn recursively. By integrating error recovery into the training loop, ReTerm significantly improves agent robustness in complex terminal environments.

LATEST/Researchers from Fudan University and others introduce ReTerm, a framework enabling code agents to recover from their own failures and learn recursively. By integrating error recovery into the training loop, ReTerm significantly improves agent robustness in complex terminal environments.

Hourly heat curve

14 fully observed hours

Now

8.8

Peak

12.012:00

24h change

-27%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

机器之心 (微信公众号)T2Nov 4, 12:00·78 pts
ReTerm Enables Recursive Self-Improvement for Code Agents via Failure Recovery
- ReTerm explores a path of Recursive Self-Improvement (RSI) driven by self-generated errors, turning failures into training data.
- It establishes a closed loop of 'encounter problem - attempt recovery - update parameters based on success/failure', allowing models to adapt to evolving failure modes.
- Experiments show that ReTerm-trained models outperform standard GRPO in accurately diagnosing environment states and finding recovery directions in cases like PyStan.
View Related Reports & Analysis Original Source