Overview
The research introduces a system named Self-Evolving Agents (SEA), an architecture designed to manage self-modification in agents. This system addresses the challenge that self-evolving agents, where the policy being updated produces its own data, evaluator, components, and hypothesis space, typically violate assumptions underlying most learning-theoretic guarantees.
At its core, SEA confines self-modification to a dedicated steering adapter and a versioned harness. These components operate around a "frozen" base model. Modifications are implemented through an "anytime-valid gate" which issues an auditable certificate against a pre-defined error budget. This gate functions by selecting among behaviors that the frozen base model can already produce.
Approach
The SEA architecture incorporates five loop controllers which compose published guarantees. To support the gate's function, five verifier-in-the-loop mechanisms provide dense, grader-free signals derived solely from the issue text. These mechanisms include:
- Best-of-$N$ selection
- Micro-step search
- Self-authored reproduction oracles
- Search-layer control
- Self-repair functionalities
Findings
The system was evaluated on a 52-instance SWE-bench Verified subset, employing four different base models. The evaluation indicated that the base capability of the models was the dominant and confound-free effect observed. On two of the strong base models, a deliberate no-op-composite control was used to isolate and measure the specific contribution of the SEA suite. This control demonstrated contributions of $+4$ and $+5$ to the performance scores. Specifically, this translated to scores changing from $24$ to $28$ for GLM 5.2 and from $29$ to $34$ for GPT, representing the 65% best performance.
Event logs were analyzed to confirm that the mechanisms within the SEA suite fired as intended and successfully prevented regressions during the self-modification processes.
Limitations Mentioned by Researchers
The results presented are derived from single-run evaluations, which are described as expensive. The researchers noted that confirming run-to-run variance and adapting the per-task algorithm mix are areas for future work.