Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces Self-Evolving Prompt Optimization (SePO), a novel approach that simultaneously optimizes both the system prompts of task agents and its own system prompt through an open-ended evolutionary search. This method allows the prompt agent to adaptively refine its instructions, leading to significant improvements in agent performance across diverse benchmarks, including math, abstract reasoning, and logic puzzles. SePO achieves an average accuracy increase of 4.49 points over existing methods, demonstrating its effectiveness in enhancing prompt-based agent behavior without altering the underlying models.
SePO not only optimizes task prompts but also evolves its own, leading to a 4.49-point accuracy boost across multiple challenging benchmarks.
System prompt optimization improves agent behavior without modifying the underlying model, yielding human-readable, model-agnostic instructions. Existing methods build a prompt agent that refines task agents'system prompts, yet leave the prompt agent's own system prompt hand-engineered and fixed. We propose Self-Evolving Prompt Optimization (SePO), which treats the prompt agent's own system prompt as an optimization target alongside task agents'system prompts. SePO adopts a self-referential design. A single prompt agent improves both task agents'system prompts and its own under an open-ended evolutionary search that maintains an archive of candidate prompts as stepping stones. Training proceeds in two stages: pre-training evolves the prompt agent on a multi-task pool, and fine-tuning then applies it to a target task. Across five benchmarks spanning math (AIME'25), abstract reasoning (ARC-AGI-1), graduate-level science (GPQA), code generation (MBPP), and logic puzzles (Sudoku), SePO consistently outperforms Manual-CoT, TextGrad, and MetaSPO, improving the average accuracy by 4.49 points compared to Manual-CoT. The prompt optimization skill from pre-training also generalizes to tasks beyond the pre-training mixture, rather than memorizing per-task prompts.