SePO: Self-Evolving Prompt Agent for System Prompt Optimization

Closing the Loop on AI Instruction

Researchers report an average accuracy improvement of 4.49 points across five benchmarks compared to manually crafted instructions. This gain comes from a new method called Self-Evolving Prompt Optimization (SePO). Most modern AI agents rely on a fixed system prompt (a set of foundational instructions) to guide their behavior. While researchers can automatically refine these instructions for a "task agent" (the worker), the "prompt agent" (the manager) responsible for the refining is usually hand-engineered by humans. This creates a bottleneck. The optimizer itself cannot learn from its mistakes or improve its teaching ability over time.

SePO breaks this ceiling by treating the prompt agent as an optimization target. Instead of a fixed instructor, SePO creates a self-referential loop. The agent responsible for writing instructions can actually improve its own instructions. This shift turns prompt optimization from a fixed tool into a learnable skill.

The Architecture of Self-Reference

At its core, SePO addresses the asymmetry in current prompt optimization. In standard setups, you have a task agent and a prompt agent. The manager looks at the worker's failures and suggests a better instruction set. However, the manager's own system prompt is written by a human and remains static. Consequently, the entire optimization process is limited by the initial quality of that human-written manual.

SePO removes this boundary through a self-referential design. The researchers propose that the prompt agent should treat itself as just another task agent. By doing so, the same evolutionary mechanism used to improve a worker's instructions can be applied to the manager's own instructions. This "closes the loop," as shown in .

As the system encounters more complex tasks, the optimizer becomes more sophisticated at diagnosing errors.

Breaking the Optimization Ceiling

To implement this, the authors employ an open-ended evolutionary search. Unlike a simple linear search, this method maintains an "archive" (a digital library) of candidate prompts. These act as stepping stones. This allows the system to explore a wider landscape of possible instructions without losing previous progress.

The authors organize this evolution into a two-stage training pipeline .

1. Stage 1: Pre-training. The prompt agent undergoes evolution across a diverse multi-task pool. During this stage, the agent improves its own system prompt. This stage cultivates a general "prompt optimization skill" rather than just memorizing specific answers.
2. Stage 2: Fine-tuning. Once the prompt agent is a "generalist" optimizer, it is deployed to a specific target task. Here, it uses its evolved expertise to refine the system prompt of a new task agent.

This two-stage approach also manages costs. For the SePO-Generalist configuration, the authors report an amortized pre-training cost of $7.43 per task. This makes the method highly efficient for deploying across many different applications.

Generalization and Performance

The effectiveness of this approach is demonstrated across five distinct benchmarks: AIME’25 (mathematics), ARC-AGI-1 (abstract reasoning), GPQA (science), MBPP (code generation), and Sudoku (logic puzzles). The study finds that SePO-Generalist consistently outperforms traditional baselines. These include manually crafted instructions (Manual-CoT), textual-gradient frameworks (TextGrad), and meta-learning methods (MetaSPO).

The researchers demonstrate that the optimization skill is truly transferable. Even when a task like Sudoku was entirely excluded from the pre-training mixture, the evolved prompt agent improved performance significantly .

This confirms the agent is not just memorizing templates. It is learning the meta-skill of how to diagnose and fix instructional failures.

Ablation studies confirm that both self-improvement and evolutionary search are vital. Removing the self-improvement step or replacing the archive-based evolution with a simple linear search resulted in lower average accuracies [Table 2].

Implications for Autonomous Agents

The success of SePO suggests a shift in how we might develop artificial intelligence. Instead of spending thousands of human hours hand-tuning instructions, we can build systems that teach themselves. By amortizing the cost of a single pre-training run, SePO offers a path toward efficient, automated deployment.

Qualitative analysis of the evolved prompts reveals how AI "thinks" about instructions. The evolved prompt agent developed "defensive principles." These include cautioning against instructions that cause an agent to truncate its reasoning. Similarly, task agents evolved granular workflows. For example, a coder might be instructed to avoid certain global-namespace collisions. These specific rules were not present in the original human-written seeds.

Limits of the Loop

Despite these advances, the framework has clear boundaries. The authors note that increasing the search depth yields diminishing returns. This suggests the system eventually hits a ceiling imposed by the underlying language models.

Additionally, the current evaluation is limited to five specific benchmarks. They do not account for more complex agentic behaviors. These include multi-turn dialogues, long-horizon planning, or the use of external tools. Whether a self-evolving agent can optimize its own ability to use a calculator remains an open question.

Figures from the paper