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

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 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 MASPO, a joint prompt optimization framework for LLM multi-agent systems that uses label-free joint rewards (local, lookahead, global), misalignment-aware sampling, and beam search with refresh under coordinate-ascent scheduling.

Problem

Prompts for agents in LLM-based multi-agent systems are hard to optimize jointly because local agent objectives misalign with global task success, and labeled data gives only sparse credit signals.

Approach

A composite joint reward that scores candidate prompts by local validity, one-step lookahead effect on successor agents, and global final-answer alignment (judged pairwise by an LLM without labels), combined with mining of misalignment cases to guide proposals, an evolutionary beam search, and a beam refresh that re-anchors scores to the current best prompt as peers change.

Key results

• MASPO on Sequential MAS reaches 70.39 average accuracy versus 65.31 for the Sequential MAS baseline and 66.56 for SPO-adapted MAS, with Hierarchical MAS MASPO reaching 71.05 average.
• MASPO reports an average gain of 2.90 over the best optimization baselines, with a 5.06 gain on Sequential MAS and 2.73 on Hierarchical MAS.
• Optimized prompts optimized on Qwen3-8B transfer to DeepSeek-V3 (71.79 to 75.86), GLM-4.6 (75.61 to 78.41), Claude-Sonnet-4 (77.58 to 79.73), and Gemini-2.5-Pro (84.93 to 87.14) average accuracy.

Surprising findings

• Prioritizing successful traces (success-case sampling) underperforms random sampling, while misalignment cases act as more informative hard negatives.
• Deeper 2-step and 3-step lookahead gave negligible gains over 1-step lookahead.

Limitations

• Global alignment and evaluation rely on Gemini-2.5-Pro as judge, and the self-optimized Qwen3-8B variant is weaker (67.70 average).
• Evaluation is limited to six benchmarks in math, graduate-level QA, and code, with a small 50-question sample pool.
• Hyperparameters (alpha, beta, theta, K, T, D) are chosen on the evaluated setup, and the reported gains are modest on some tasks.

Experimental setup

Datasets: MATH-500 AQuA AGIEval-MATH GPQA-Diamond MBPP HumanEval-ET

Benchmarks: MATH-500 AGIEval-MATH AQuA GPQA MBPP HumanEval-ET

Metrics: accuracy

Baselines: Vanilla CoT SC (CoT) Self-Refine AgentDropout TPE SPO Sequential MAS Hierarchical MAS

Builds on: SPO MIPRO MASS AgentDropout

Closest papers by meaning

• MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks (ICML 2026 · similarity 0.91)
• Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.91)
• Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.89)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.89)
• AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (NeurIPS 2026 · similarity 0.89)
• MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems (ICML 2025 · similarity 0.89)