SAGE Improves Autonomous AI Research by Self-Correcting Experimental Failures.

Autonomous research agents are becoming more sophisticated, capable of generating hypotheses, coding, running experiments, and even drafting papers. However, a major challenge remains in their ability to recover effectively when experiments fail. Current methods often rely on a single, broad reflection, which can lead to inefficient trial-and-error or discarding valuable context. A new system called SAGE (Self-correcting, Autonomous, Grounded Experimenter) addresses this by introducing Multi-Hypothesis Failure Attribution (MHFA). MHFA treats failure recovery as a structured causal diagnosis, generating multiple evidence-based explanations for a failure, evaluating their severity, and routing the root cause to the appropriate intervention level (hypothesis, experimental design, or implementation). SAGE also incorporates a grounded reporting mechanism to ensure scientific integrity by constraining drafted results to actual measured values, preventing AI hallucinations. Benchmarking shows SAGE dramatically increases metrics-bearing outputs and improves artifact quality compared to reflection-based baselines, making it a more trustworthy foundation for future autonomous research.