Research map / Multi-agent systems and games / LLM agents

SePO: Self-Evolving Prompt Agent for System Prompt Optimization

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SePO, a self-referential prompt optimization framework where a prompt agent evolves its own system prompt alongside task agents' prompts via archive-based open-ended evolution, in two stages (pre-training on a task pool, then fine-tuning on a target task).

Problem

Existing system prompt optimization methods keep the prompt agent's own system prompt hand-engineered and fixed, bounding optimization by what humans can write and preventing accumulation of experience across tasks.

Approach

The prompt agent is treated as a special task agent so the same archive-based open-ended evolutionary procedure refines both its own system prompt and task agents' prompts; pre-training evolves the prompt agent on a multi-task pool and fine-tuning reuses it on a target task.

Key results

• SePO-Generalist reaches 76.38 average accuracy vs 71.89 for Manual-CoT (+4.49), with the best score on all five tasks.
• On ARC-AGI-1 SePO-Generalist scores 43.39 vs 37.30 for Manual-CoT; on Sudoku (held out of pre-training) it reaches 99.90 vs 96.95.
• Under a model swap (Gemini 3.1 Flash-Lite task agent, Claude Opus 4.6 prompt agent), average accuracy improves from 67.95 to 70.08.

Surprising findings

• Removing open-ended evolution (linear search) drops AIME'25 accuracy the most (-6.98 points), though it still improves ARC-AGI-1 over the full method's ablation.
• TextGrad and MetaSPO fall below the Manual-CoT baseline on average (70.39 and 71.32 vs 71.89).

Limitations

• Gains are modest on several tasks (e.g., GPQA +1.7 and AIME'25 +6.7 over Manual-CoT in the default setting), and results rely on a single run configuration averaged over 5 seeds.
• Multi-task pool selection uses a greedy heuristic with an LLM judge; more advanced task selection is left for future work.
• Relies on proprietary frontier models (Gemini 3.1 Pro Preview, DeepSeek-V3.2) with no open-weight results.

Experimental setup

Datasets: AIME'25 ARC-AGI-1 GPQA MBPP Sudoku simplescaling/s1K-1.1 GAIR/LIMO-v2 MMLU

Benchmarks: AIME'25 ARC-AGI-1 GPQA MBPP Sudoku

Metrics: accuracy pass@1 pass@3

Baselines: Manual-CoT TextGrad MetaSPO SePO-Specialist

Builds on: PromptBreeder TextGrad MetaSPO Darwin Gödel Machine ADAS

Closest papers by meaning

• PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization (ICLR 2024 · similarity 0.89)
• Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution (ICML 2024 · similarity 0.89)
• MARS: Multi-Agent Adaptive Reasoning with Socratic Guidance for Automated Prompt Optimization (AAAI 2026 · similarity 0.88)
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.88)
• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.88)
• PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.88)